The DevSecOps Talks Podcast

We are always happy to answer any questions, hear suggestions for new episodes, or hear from you, our listeners.

DevSecOps Talks podcast LinkedIn page

DevSecOps Talks podcast website

DevSecOps Talks podcast YouTube channel

In this episode, Mattias, Andre, and Paulina welcome back returning guest Paul from System Initiative to continue a conversation that started in the previous episode about their project Swamp. The discussion digs into how AI-assisted software development has changed over the past year, and why the real shift is not "AI writes code" but humans orchestrating multiple specialized agents with strong guardrails. Paul walks through the practical workflows, multi-layered testing, architecture-first thinking, cost discipline, and security practices his team has adopted — while the hosts push on how this applies across enterprise environments, mentoring newcomers, and the uncomfortable question of who is responsible when AI-built software fails.

Before diving into technical specifics, Paul acknowledges that the industry is at "a real crazy crossroads." He references Block (formerly Square) cutting roughly 40% of their workforce, citing uncertainty about what AI means for their teams. He wants to be transparent that System Initiative also shrank — but clarifies the company did not cut people because of AI. The decision to reduce headcount came before they even knew what they were going to build next, let alone how they would build it. AI entered the picture only after they started prototyping the next version of their product.

Block's February 2026 layoffs, announced by CEO Jack Dorsey, eliminated over 4,000 positions. The move was framed as an AI-driven restructuring, making it one of the most visible examples of AI anxiety playing out in real corporate decisions.

Paul explains that AI coding quality shifted significantly around October–November of the previous year. Before that, results were inconsistent — sometimes impressive, often garbage. Then the models improved dramatically in both reasoning and code generation.

But the bigger breakthrough, in his view, was not the models themselves. It was the industry's shift from "Gen AI" — one-shot prompting where you hand over a spec and accept whatever comes back — to agentic AI, where the model acts more like a pair programmer. In that setup, the human stays in the loop, challenges the plan, adds constraints, and steers the result toward something that fits the codebase.

He gives a concrete early example: System Initiative had a CLI written in Deno (a TypeScript runtime). Because the models were well-trained on TypeScript libraries and the Deno ecosystem, they started producing decent code. Not beautiful, not perfectly architected — but functional. When Paul began feeding the agent patterns, conventions, and existing code to follow, the output became coherent with their codebase.

This led to a workflow where Paul would open six Claude Code sessions at once in separate Git worktrees — isolated copies of the repository on different branches — each building a small feature in parallel, feeding them bug reports and data, and continuously interacting with the results rather than one-shotting them.

Git worktrees let you check out multiple branches of the same repository simultaneously in separate directories. Each worktree is independent, so you can work on several features at once and merge them back via pull requests.

He later expanded this by running longer tasks on a Mac Mini accessible via Tailscale (a mesh VPN), while handling shorter tasks on his laptop — effectively distributing AI workloads across machines.

One of Paul's strongest themes is that AI shifts engineering attention away from syntax and back toward architecture. He argues that AI can generate plenty of code, but without design principles and boundaries it will produce spaghetti on top of existing spaghetti.

He introduces the idea of "the first thousand lines" — an anecdote he read recently claiming that the first thousand lines of code an agent helps write determine its path forward. If those lines are well-structured and follow clear design principles, the agent will build coherently on top of them. If they are messy and unprincipled, everything after will compound the mess.

Paul breaks software development into three layers:

He argues the industry spent the last decade obsessing over "taste" while often mocking "ivory tower architects" — the people who designed systems but didn't write code. In an AI-driven world, those architectural concerns become critical again because the agent needs clear boundaries, domain structure, and intent to produce coherent output.

Paulina agrees and observes that this trend may also blur traditional specialization lines, pushing engineers toward becoming more general "software people" rather than narrowly front-end, back-end, or DevOps specialists.

Paul describes how his team makes architecture actionable for AI by encoding system knowledge directly into the repository. Their approach has several layers:

Design documents — Detailed docs covering the model layer (the actual objects, their purposes, how they relate), workflow construction (how models connect and pass data), and expression language behavior. These live in a /design folder in the open-source repo and describe the intent of every part of the system.

Architectural rules — The agent is explicitly told to follow Domain-Driven Design: proper separation between domains, infrastructure, repositories, and output layers. The DDD skill is loaded so the agent understands and maintains bounded contexts.

Code standards — TypeScript strict mode, no any types, named exports, passing lint and format checks. License compliance is also enforced: because the project is AGPL v3, the agent cannot pull in dependencies with incompatible licenses.

Skills — A newer mechanism for lazy-loading contextual information into the AI agent. Rather than stuffing everything into one enormous prompt, skills are loaded on demand when the agent encounters a specific type of task. This keeps context windows lean and focused.

AGPL v3 (GNU Affero General Public License) is a copyleft license that requires anyone who runs modified software over a network to make the source code available. This creates strict constraints on what dependencies can be used.

A major part of the discussion centers on how Paul's team works with multiple specialized AI agents rather than a single all-knowing assistant. The chain looks like this:

Issue triage agent — When a user opens a GitHub issue, an agent evaluates whether it is a legitimate feature request or bug report. The agent's summary is posted back to the issue immediately, creating context for later stages.

Planning agent — If the issue is legitimate, the system enters plan mode. A specification is generated and posted for the user to review. Users can push back ("that's not how I think it should work"), and the plan is revised until everyone agrees.

Implementation agent — The code is written based on the approved plan, with all the design docs, architectural rules, and skills loaded as context.

Happy-path reviewer — A separate agent reviews the code against standards, checking that it loads correctly and appears to function.

Adversarial reviewer — Added just days before the recording, this agent is told: "You are a grumpy DevOps engineer and I want you to pull this code apart." It looks for security injection points, failure modes, and anything the happy-path reviewer might miss.

Both review agents write their findings as comments on the pull request, creating a visible audit trail. The PR only merges when both agents approve. If the adversarial agent flags a security vulnerability, the implementation goes back for changes.

Paul says this "Jekyll and Hyde" review setup caught a path traversal bug in their CLI during its first week. While the CLI runs locally and the risk was limited, it proved the value of adversarial review.

Path traversal is a vulnerability where an attacker can access files outside the intended directory by manipulating file paths (e.g., using ../ sequences). Even in CLI tools, this can expose sensitive files on a user's machine.

Mattias compares the overall process to a modernized CI/CD pipeline — the same stages exist (commit, test, review, promote, release), but AI replaces some of the manual implementation steps while humans stay focused on architecture, review, and acceptance.

One of the more provocative decisions Paul describes: the open-source Swamp project does not accept external pull requests. GitHub recently added a feature to disable PR creation from non-collaborators entirely, and the team turned it on immediately.

The reasoning is supply chain control. Because the project's code is 100% AI-generated within a tightly controlled context — design docs, architectural rules, skills, adversarial review — they want to ensure that all code entering the system passes through the same pipeline. External PRs would bypass that chain of custody.

Contributors are instead directed to open issues. The team will work through the design collaboratively, plan it together, and then have their agents implement it. Paul frames this not as rejecting collaboration but as controlling the process: "We love contributions, but in the AI world, we cannot control where that code is from or what that code is doing."

The team built a skill into Swamp itself so that when the tool encounters a bug during use, it can check out the version of the source code the binary was built against, analyze the problem, and open a GitHub issue automatically with detailed context.

This creates high-quality bug reports that already contain the information needed to reason about a fix. When the implementation agent later picks up that issue, it has precise context — where the bug is, what triggered it, and what the expected behavior should be. Paul says the quality of issues generated this way is significantly higher than typical user-filed bugs.

Although the conversation starts with code generation, Paul says testing is actually his favorite part of the workflow. The team runs multiple layers:

Product-level tests:

Architectural fitness tests are automated checks that verify a system's structure conforms to its intended architecture. In DDD, this means ensuring bounded contexts don't leak dependencies across domain boundaries.

User-level tests (separate repo):

Adversarial tests (multiple tiers):

Only after all these layers pass does a build get promoted from nightly to stable. Paul can download and manually test any nightly build that maps back to a specific commit.

Paulina points out that if AI is a force multiplier, there is now even less excuse not to write tests. Paul agrees: "We were scraping the barrel before at coming up with reasons why there shouldn't be any tests. Now that's eliminated."

Paul repeatedly emphasizes "plan mode," particularly in Claude Code. Before the agent changes anything, it produces a detailed plan describing what it intends to do and why, and waits for human approval.

The hosts immediately draw a parallel to terraform plan — the value is not just automation, but the chance to inspect intended changes before applying them. Paul says this was one of the biggest improvements in AI-assisted development because it reduces horror-story scenarios where an agent goes off and deletes a database or rewrites an application.

He notes that other tools are starting to adopt plan mode because it produces better results across the board. But he also warns that plan mode only helps if people actually read the plan — just like Terraform, the safeguard depends on human discipline. "If there's a big line in the middle that says 'I'm going to delete a database' and you haven't read it — it's the same thing."

Paul shares several tactical lessons:

He also notes that the generated TypeScript is not always how a human would write it — but that matters less if the result is well-tested, secure, and respects domain boundaries. "The actual syntax of the code itself can change 12 times a day. It doesn't really matter as long as it adheres to the product."

Despite all the automation, Paul is adamant that humans remain involved at every stage. Plans are reviewed, implementations are questioned, pull request comments are inspected, binaries are tested before reaching stable release. He describes it as "continually interacting with Claude Code" rather than just letting things happen.

When Paulina pushes on whether a human still checks things before production, Paul makes clear: yes, always. The release pipeline goes from commit to nightly to manual verification to stable promotion. "I will always download something before it goes to stable."

Paul acknowledges that running multiple agents in parallel is not for everyone. Context switching has always been expensive for engineers, and commanding multiple agents simultaneously is a new form of it. His advice: if you work best focusing on a single task, don't force the multi-agent style. "It'll be such a context switching killer and it'll cause you to lose focus."

The key shift is that instead of writing code, you are "commanding architecture and commanding design." But that still requires focus and judgment.

Paulina captures the dynamic bluntly: "It's a multiplier. If there is a good thing, you'll get a lot of good thing. If it's a shit, you're going to get a lot of shit."

Paul argues that experienced software and operations people are still essential because they understand architecture, security, constraints, and tradeoffs. AI amplifies whatever is already there — good engineering or bad engineering alike.

He believes engineers who learn to use these tools well become "even more important to your company than you already are." But he also acknowledges that some people will not want to work this way, and that friction between AI-forward and AI-resistant teams is already happening in organizations.

Paulina raises this personally — she was recently asked to mentor someone entering IT and struggled with how to approach it. She doesn't have a formal IT education (she has an engineering background) and learned on the go. The skills she built through manual work — understanding when code needs refactoring as scale changes, knowing how to structure projects at different sizes — are hard to teach when AI handles so much of the writing.

Paul agrees this is an open question and says the industry is still figuring out the patterns. He believes teaching principles, architecture, and core engineering fundamentals becomes even more important, because tool-specific syntax is increasingly handled by AI. "Do you need to know how to write a Terraform module? Do you need to know how to write a Pulumi provider?" — these are becoming less essential as individual skills, while understanding how systems fit together matters more.

He frames this as an opportunity: "We are now in control of helping shape how this moves forward in the industry." As innovators and early adopters, current practitioners can set the patterns. If they don't, someone else will.

Paulina raises a concrete example from Poland: someone built an app using AI to upload receipts to a centralized accounting system, released it publicly, and exposed all their customers' data.

This leads to a deeper question from Mattias about responsibility: if someone with no engineering background builds an insecure app using an AI tool, who is accountable? The user? The platform? The model provider? The episode doesn't settle this, but Paul argues it reinforces why skilled engineers remain essential. The AI doesn't know the security boundary unless someone explicitly teaches it — "it probably wasn't fed that information that it had to think about the security context."

He expects more specialized skills and agents focused on security, accessibility, and compliance to emerge — calling out the example of loading a security skill and an accessibility skill when you know an app will be public-facing. But he says the ecosystem is not fully there yet.

Paul addresses economics directly. His five-person team at System Initiative all use Claude Max Pro at $200 per person per month. They do not exceed that cost for the full AI workflow — code generation, reviews, planning, and adversarial testing.

In contrast, he has seen other organizations spend $10,000–$12,000 per month per developer on AI tokens because they let tools roam with huge context windows and vague instructions. His conclusion: tightly scoped tasks are not just better for quality — they are far cheaper.

This maps directly to classic engineering wisdom. Tightly defined stories and tasks were always more efficient to push through a system than "go rebuild this thing and I'll see you in six months." The same principle applies to AI agents.

For teams in companies that ban or restrict AI, Paul suggests a pragmatic entry point: use agents to analyze, not to write code.

He describes a conversation with someone in London who asked how to get started. Paul's advice: if you already know roughly where a bug lives, ask the agent to analyze the same bug report. If it identifies the same area and the same root cause, you have evidence that the tool can accelerate diagnosis. Show your CTO: "I'm diagnosing bugs 50% faster with this agent. It's not writing code — it's helping me understand where the issue is."

Similar analysis-first use cases work for accessibility reviews, security scans, or code quality assessments. The point is to build trust before expanding scope. Paul notes this approach works faster in the private sector than the public sector, where technology adoption has always been slower.

Paul believes the conversation has shifted dramatically in the past six months — from AI horror stories and commiserating over drinks to genuine success stories and conferences forming around agentic engineering practices. He points to two upcoming events:

His prediction: the pace is not linear. "We're honestly exponential at this moment in time." He sidesteps the ethics of AI companies (referencing tensions between Anthropic and OpenAI) to focus on the practical reality that models, reasoning, and tooling are all improving at a compounding rate.

Andrey and Mattias share a fast re:Invent roundup focused on AWS security. What do VPC Encryption Controls, post-quantum TLS, and org-level S3 block public access change for you? Which features should you switch on now, like ECR image signing, JWT checks at ALB, and air-gapped AWS Backup? Want simple wins you can use today?

We are always happy to answer any questions, hear suggestions for new episodes, or hear from you, our listeners.

DevSecOps Talks podcast LinkedIn page

DevSecOps Talks podcast website

DevSecOps Talks podcast YouTube channel

In this episode, Andrey and Mattias deliver a security-heavy recap of AWS re:Invent 2025 announcements, while noting that Paulina is absent and wishing her a speedy recovery. Out of the 500+ releases surrounding re:Invent, they narrow the list down to roughly 20 features that security-conscious teams can act on today — covering encryption, access control, detection, backups, container security, and organization-wide guardrails. Along the way, Andrey reveals a new AI-powered product called Boris that watches the AWS release firehose so you don't have to.

The hosts frame the episode as the DevSecOps Talks version of a re:Invent recap, complementing a FivexL webinar held the previous month. Despite the podcast's name covering development, security, and operations, the selected announcements lean heavily toward security. Andrey is upfront about it: if security is your thing, stay tuned; otherwise, manage your expectations.

At the FivexL webinar, attendees were asked to prioritize areas of interest across compute, security, and networking. AI dominated the conversation, and people were also curious about Amazon S3 Vectors — a new S3 storage class purpose-built for vector embeddings used in RAG (Retrieval-Augmented Generation) architectures that power LLM applications. It is cost-efficient but lacks hybrid search at this stage.

One of the first and most praised announcements is VPC Encryption Control for Amazon VPC, a pre-re:Invent release that lets teams audit and enforce encryption in transit within and across VPCs. The hosts highlight how painful it used to be to verify internal traffic encryption — typically requiring traffic mirroring, spinning up instances, and inspecting packets with tools like Wireshark. This feature offers two modes: monitor mode to audit encryption status via VPC flow logs, and enforce mode to block unencrypted resources from attaching to the VPC.

Mattias adds that compliance expectations are expanding. It used to be enough to encrypt traffic over public endpoints, but the bar is moving toward encryption everywhere, including inside the VPC. The hosts also call out a common pattern: offloading SSL at the load balancer and leaving traffic to targets unencrypted. VPC encryption control helps catch exactly this kind of blind spot.

The discussion then shifts to post-quantum cryptography (PQC) support rolling out across AWS services including S3, ALB, NLB, AWS Private CA, KMS, ACM, and Secrets Manager. AWS now supports ML-KEM (Module Lattice-Based Key Encapsulation Mechanism), a NIST-standardized post-quantum algorithm, along with ML-DSA (Module Lattice-Based Digital Signature Algorithm) for Private CA certificates.

The rationale: state-level actors are already recording encrypted traffic today in a "harvest now, decrypt later" strategy, betting that future quantum computers will crack current encryption. Andrey notes that operational quantum computing feels closer than ever, making it worthwhile to enable post-quantum protections now — especially for sensitive data traversing public networks.

Several S3-related updates stand out. Attribute-Based Access Control (ABAC) for S3 allows access decisions based on resource tags rather than only enumerating specific actions in policies. This is a powerful way to scope permissions — for example, granting access to all buckets tagged with a specific project — though it must be enabled on a per-bucket basis, which the hosts note is a drawback even if necessary to avoid breaking existing security models.

The bigger crowd-pleaser is S3 Block Public Access at the organization level. Previously available at the bucket and account level, this control can now be applied across an entire AWS Organization. The hosts call it well overdue and present it as the ultimate "turn it on and forget it" control: in 2026, there is no good reason to have a public S3 bucket.

Amazon ECR Managed Image Signing is a welcome addition. ECR now provides a managed service for signing container images, leveraging AWS Signer for key management and certificate lifecycle. Once configured with a signing rule, ECR automatically signs images as they are pushed. This eliminates the operational overhead of setting up and maintaining container image signing infrastructure — previously a significant barrier for teams wanting to verify image provenance in their supply chains.

AWS Backup gets significant attention. The hosts discuss air-gapped AWS Backup Vault support as a primary backup target, positioning it as especially relevant for teams where ransomware is on the threat list. These logically air-gapped vaults live in an Amazon-owned account and are locked by default with a compliance vault lock to ensure immutability.

The strong recommendation: enable AWS Backup for any important data, and keep backups isolated in a separate account from your workloads. If an attacker compromises your production account, they should not be able to reach your recovery copies. Related updates include KMS customer-managed key support for air-gapped vaults for better encryption flexibility, and GuardDuty Malware Protection for AWS Backup, which can scan backup artifacts for malware before restoration.

Dynamic data masking in Aurora PostgreSQL draws praise from both hosts. Using the new pg_columnmask extension, teams can configure column-level masking policies so that queries return masked data instead of actual values — for example, replacing credit card numbers with wildcards. The data in the database remains unmodified; masking happens at query time based on user roles.

Mattias compares it to capabilities already present in databases like Snowflake and highlights how useful it is when sharing data with external partners or other teams. When the idea of using masked production data for testing comes up, the hosts gently push back — don't do that — but both agree that masking at the database layer is a strong control because it reduces the risk of accidental data exposure through APIs or front-end applications.

The episode covers several IAM-related features. AWS IAM Outbound Identity Federation allows federating AWS identities to external services via JWT, effectively letting you use AWS identity as a platform for authenticating to third-party services — similar to how you connect GitHub or other services to AWS today, but in the other direction.

The AWS Login CLI command provides short-lived credentials for IAM users who don't have AWS IAM Identity Center (SSO) configured. The hosts see it as a better alternative than storing static IAM credentials locally, but also question whether teams should still be relying on IAM users at all — their recommendation is to set up IAM Identity Center and move on.

The AWS Source VPC ARN condition key gets particular enthusiasm. It allows IAM policies to check which VPC a request originated from, enabling conditions like "allow this action only if the request comes from this VPC." For teams doing attribute-based access control in IAM, this is a significant addition.

AWS Secrets Manager Managed External Secrets is another useful feature that removes a common operational burden. Previously, rotating third-party SaaS credentials required writing and maintaining custom Lambda functions. Managed external secrets provides built-in rotation for partner integrations — Salesforce, BigID, and Snowflake at launch — with no Lambda functions needed.

JWT verification in AWS Application Load Balancer simplifies machine-to-machine and service-to-service authentication. Teams previously had to roll their own Lambda-based JWT verification; now it is supported out of the box. The recommendation is straightforward: drop the Lambda and use the built-in capability.

AWS Network Firewall Proxy is in public preview. While the hosts have not explored it deeply, their read is that it could help with more advanced network inspection scenarios — not just outgoing internet traffic through NAT gateways, but potentially also traffic heading toward internal corporate data centers.

Although the episode is mainly security-focused, the hosts include REST API streaming in Amazon API Gateway as a nod to developers. This enables progressive response payload streaming, which is especially relevant for LLM use cases where streaming tokens to clients is the expected interaction pattern. Mattias notes that applications are moving beyond small JSON payloads — streaming is becoming table stakes as data volumes grow.

CloudWatch unified management for operational, security, and compliance data promises cross-account visibility from a single pane of glass, without requiring custom log aggregation pipelines built from Lambdas and glue code. The hosts are optimistic but immediately flag the cost: CloudWatch data ingest pricing can escalate quickly when dealing with high-volume sources like access logs. Deep pockets may be required.

Detection is a recurring theme throughout the episode. The hosts discuss CloudTrail Insights for data events (useful if you are already logging data-plane events — another deep-pockets feature), extended threat detection for EC2 and ECS in GuardDuty using AI-powered analysis to correlate security signals across network activity, runtime behavior, and API calls, and the public preview of AWS Security Agent for automated security investigation.

On GuardDuty specifically, the recommendation is clear: if you don't have it enabled, go enable it — it gives you a good baseline that works out of the box across your services with minimal setup. You can always graduate to more sophisticated tooling later, but GuardDuty is the gap-stopper you start with.

Mattias drives the broader point home: incidents are inevitable, and what you can control is how fast you detect and respond. AWS is clearly investing heavily in the detection side, and teams should enable these capabilities as fast as possible.

Several updates make governance easier to adopt at scale: - Dedicated controls for AWS Control Tower without requiring a full Control Tower deployment — you can now use Control Tower guardrails à la carte. - Automatic enrollment in Control Tower — a feature the hosts feel should have existed already. - Required tags in Organizations stack policies — enforcing tagging standards at the organization level. - Amazon Inspector organization-wide management — centralized vulnerability scanning across all accounts. - Billing transfer for AWS Organizations — useful for AWS resellers managing multiple organizations. - Delete protection for CloudWatch Log Groups — a small but important safeguard.

Mattias says plainly: everyone should use Control Tower.

The conversation shifts to the public preview of AWS MCP (Model Context Protocol) servers. Unlike traditional locally-hosted MCP servers that proxy LLM requests to API calls, AWS is taking a different approach with remote, fully managed MCP servers hosted on AWS infrastructure. These allow AI agents and AI-native IDEs to interact with AWS services over HTTPS without running anything locally.

AWS launched four managed MCP servers — AWS, EKS, ECS, and SageMaker — that consolidate capabilities like AWS documentation access, API execution across 15,000+ AWS APIs, and pre-built agent workflows. However, the IAM model is still being worked out: you currently need separate permissions to call the MCP server and to perform the underlying AWS actions it invokes. The hosts treat this as interesting but still evolving.

Toward the end of the episode, Andrey reveals a personal project: Boris (getboris.ai), an AI-powered DevOps teammate he has been building. Boris connects to the systems an engineering team already uses and provides evidence-based answers and operational automation.

The specific feature Andrey has been working on takes the AWS RSS feed — where new announcements land daily — and cross-references it against what a customer actually has running in their AWS Organization. Instead of manually sifting through hundreds of releases, Boris sends a digest highlighting only the announcements relevant to your environment and explaining how you would benefit.

Mattias immediately connects this to the same problem in security: teams are overwhelmed by the constant flow of feature updates and vulnerability news. Having an AI that filters and contextualizes that information is, in his words, "brilliant."

Andrey also announces that Boris has been accepted into the Tehnopol AI Accelerator in Tallinn, Estonia — a program run by the Tehnopol Science and Business Park that supports early-stage AI startups — selected from more than 100 companies.

We are always happy to answer any questions, hear suggestions for new episodes, or hear from you, our listeners.

DevSecOps Talks podcast LinkedIn page

DevSecOps Talks podcast website

DevSecOps Talks podcast YouTube channel

System Initiative has undergone a dramatic transformation: from a visual SaaS infrastructure platform with 17 employees to Swamp, a fully open-source CLI built for AI agents, maintained by a five-person team whose initials literally spell the product name. Paul Stack returns for his third appearance on the show to explain why the old model failed — and why handing an AI agent raw CLI access to your cloud is, as Andrey puts it, just "console-clicking in the terminal." The conversation gets sharp when the hosts push on what problem Swamp actually solves, whether ops teams are becoming the next bottleneck in AI-era delivery, and why Paul believes the right move is not replacing Terraform but giving AI a structured system it can reason about. Paul also drops a parting bombshell: he hasn't written a single line of code in four weeks.

Paul Stack explains that System Initiative spent over five years iterating on a visual infrastructure tool where users could drag, drop, and connect systems. Despite the ambition, the team eventually concluded that visual composition was too slow, too cumbersome, and too alien for practitioners accustomed to code, artifacts, and reviewable changes.

The shift started in summer 2025 when Paul spiked a public OpenAPI-spec API. A customer then built an early MCP (Model Context Protocol) server on top of it — a prototype that worked but had no thought given to token usage or tool abstraction. System Initiative responded by building its own official MCP server and pairing it with a CLI. The results were dramatically better: customers could iterate easily from the command line or through AI coding tools like Claude Code.

By Christmas 2025 the writing was on the wall. The CLI-plus-agent approach was producing better outcomes, while the company was still carrying hundreds of thousands of lines of code for a distributed SaaS platform built for a previous product direction. In mid-January 2026, the company made the call to rethink everything from first principles.

The restructuring was painful. System Initiative went from 17 people to five. Paul explains the reasoning candidly: when you don't know what the tool is going to be, keeping a large team around is unfair to them, bad for their careers, and expensive. The five who stayed were the CEO, VP of Business, COO, Paul (who ran product), and Nick Steinmetz, the head of infrastructure — who also happened to be System Initiative's most active internal user, having used the platform to build System Initiative itself.

Those five people's initials spell SWAMP. The name was unintentional but stuck — and Paul notes with a grin that if they ever remove the "P," it becomes "SWAM," so he's safe even if he leaves. Beyond the joke, the name fits: Swamp stores operational data in a local .swamp/ directory — not a neatly formatted data lake, but a structured store that AI agents can pull from to reason about infrastructure state and history.

A major theme in the conversation is that letting an AI agent operate infrastructure directly — through the AWS CLI or raw API calls — is fundamentally unreliable. Andrey lays out the problem clearly: this kind of interaction is equivalent to clicking around the cloud console, just automated through a terminal. It is not repeatable, not reviewable, and inherits the non-deterministic behavior of LLMs. If the agent's context window fills up, it starts to forget earlier decisions and improvises — a terrifying prospect for production infrastructure.

What made System Initiative's earlier MCP-based direction compelling, in Andrey's view, was the combination of guardrails, repeatability, and human review. The agent generates a structured specification, a human reviews it, and only then is it applied. Paul agrees and calls this the "agentic loop with the human loop" — the strongest pattern they found.

Paul shares a hard-won lesson from building MCP integrations: a poorly designed MCP server burns enormous amounts of tokens and creates unnecessary costs for users. He spent three weeks in December reworking the server to use progressive context reveal rather than flooding the model with data. Even so, the fundamental problem with a SaaS-first architecture remained — constantly transmitting context between a central API and the user's agent was expensive regardless of optimization.

That experience pushed the team toward a local-first design. Swamp keeps data on the user's machine, close to where the agent operates, giving AI the context it needs without the round-trip overhead and cost of a remote service.

Swamp is a general-purpose, open-source CLI automation tool — not just another infrastructure-as-code framework. Its core building blocks are:

Critically, Swamp ships with zero built-in models — no pre-packaged AWS EC2, VPC, or GCP resource definitions. Instead, the AI agent uses installed skills to generate models on the fly. Paul describes a user who joined the Discord that very morning, asked Swamp to create a schema for managing Let's Encrypt certificates, and it worked on the first attempt without writing any code.

Nick Steinmetz provides another example: he manages his homelab Proxmox hypervisor entirely through Swamp — creating and starting VMs, inspecting hypervisor state, and monitoring utilization. He recently connected it to Discord so friends can run commands like @swamp create vm to spin up Minecraft and gaming servers on demand.

The hosts spend significant time pinning down where Swamp sits relative to tools like Claude Code, bash access, and existing automation. Paul is clear: Swamp is not an AI wrapper or chatbot. It is a structured runtime that gives agents guardrails and reusable patterns.

Mattias works through several analogies to help frame it — is it like n8n or Zapier for the CLI? A CLI-based Jenkins where jobs are agents? Paul settles on this: it is a workflow engine driven by typed models, where data can be chained between steps using CEL (Common Expression Language) expressions — the same dot-notation referencing used in Kubernetes API declarations. A simple example: create a VPC in step one, then reference VPC.resource.attributes.vpcid as input to a subnet model in step two.

In Paul's personal workflow, he uses Claude Code to generate models and workflows, checks them into Git for peer review, and then runs them manually or through CI at a time of his choosing. He has explicitly configured Claude with a permission deny on workflow run — the agent helps build automation but never executes it. The same CLI works whether a person or an agent runs it; the difference is timing and approval.

Swamp workflows are parameterized and reusable across environments. If they grow unwieldy, workflows can orchestrate other workflows, collect outputs, and manage success conditions — similar to GitHub Actions calling other actions.

Paul also demonstrates that Swamp can sit alongside existing tooling rather than replacing it. In a live Discord session, he built infrastructure models in Swamp and then asked the AI agent to generate the equivalent Terraform configuration. Because the agent had typed models with explicit relationships, it produced correct Terraform with proper resource dependencies. This positions Swamp less as a replacement mandate and more as a reasoning and control layer that can output to whatever format teams already use.

When one of the hosts compares Swamp to general build systems like Gradle, Paul draws a key distinction: traditional tools were designed for humans to write, review, and debate. Swamp is designed for AI agents to inspect and operate within. He references Anton Babenko's widely-used terraform-aws-vpc module — with its 237+ input variables — as an example of a human-centric design that agents struggle with due to version dependencies, module structure complexity, and stylistic decisions baked in over years. Swamp instead provides the agent with structured context, explicit typing, and historical artifacts it can query.

Paulina asks the natural question: if Swamp is fully open source under AGPL v3, how does the company make money?

Paul is candid that monetization is not the immediate priority — the focus is building a tool that resonates with users first. But he outlines a potential model: a marketplace-style ecosystem where users can publish their own models and workflows, while System Initiative offers supported, maintained, and paid-for versions of commonly needed building blocks. He draws a loose comparison to Docker Hub's model of community images alongside official ones.

The deeper argument is strategic: Paul believes there is no longer a durable moat in software. If users dislike a tool today, AI makes it increasingly feasible to build their own. Rather than trying to control all schemas and code, the team wants to make Swamp so extensible that users build on top of it rather than walking away from it.

Paul argues that software development productivity is accelerating so fast with AI that ops teams risk becoming the next bottleneck — echoing earlier industry transitions from physical servers to cloud and from manual provisioning to infrastructure as code. Development teams can now move at a pace that traditional infrastructure workflows cannot match.

Andrey agrees with the premise but pushes back on where the bottleneck actually sits today. In his experience — spending "day and night burning tokens" on AI-assisted development — the real constraint is testing, not deployment. He describes pipelines that can go from idea to pull request automatically, but stall without a strong test harness and end-to-end validation. Without sufficient tests, you never even reach the deployment phase.

Paul accepts the framing and says the goal of Swamp is to strip away lower-value friction — fighting with file layouts, naming conventions, writing boilerplate models — so teams can invest their time where engineering rigor still matters most: testing, validation, and production safety.

Paul closes with an important positioning point: Swamp does not require teams to discard their Terraform, Pulumi, or existing infrastructure investments. It can be introduced alongside current tooling to interrogate infrastructure, validate what existing IaC does, and extend automation in AI-native ways. The extensibility is the point — users control when things run, what models to build, and how to integrate with their existing stack.

Andrey challenges the assumption that AI-driven infrastructure is automatically safer. If an agent is just shelling out to the AWS CLI, the result may be fast — but it is non-deterministic, non-repeatable, and forget-prone once the context window fills up.

The future of infra automation needs guardrails before it needs speed. Listen to hear why structured workflows beat flashy demos.

The breakthrough was not autonomous infrastructure execution. It was letting the AI generate structured specs while humans stay in charge of review and execution. Paul even blocks Claude Code from running workflows directly on his machine.

If "human in the loop" sounds conservative, this episode makes the case that it is the only production-safe pattern we have. Listen for the full argument.

Paul argues that AI has changed the economics of building software so fundamentally that no team can rely on implementation complexity as a competitive advantage. If users dislike your tool, they can build their own — faster than ever before.

That belief is why Swamp is open source, extensible, and ships with zero built-in models. Listen for a candid take on product strategy when anyone can clone your work.

As development velocity explodes with AI, Paul warns that infrastructure teams risk slowing everything down — the same pattern that played out in the shifts from physical servers to cloud and from cloud to IaC.

Andrey fires back: the real bottleneck today is testing, not deployment. Listen for a sharp debate on where delivery pipelines are actually stuck.

Paul reveals that the entire Swamp repository is AI-generated, with four machines running in parallel to churn out plans and implementations — including customer feature requests. The team teases a future episode to compare notes on AI-driven development workflows.

If that claim doesn't make you want to hear the follow-up, nothing will.

Swamp CLI on GitHub — The open-source, AGPL v3 licensed CLI tool discussed in the episode. Models, workflows, and a local .swamp/ data directory designed for AI agent interaction.

System Initiative — The company behind Swamp, originally known for its visual infrastructure platform, now pivoted to AI-native CLI automation.

Model Context Protocol (MCP) — Anthropic's open protocol for connecting AI models to external tools and data sources. Paul discusses the challenges of building MCP servers that are token-efficient.

Claude Code — Anthropic's agentic coding tool that runs in the terminal. Used throughout the episode as the primary AI agent interface for Swamp workflows.

CEL — Common Expression Language — The expression language Swamp uses for chaining data between workflow steps, similar to how Kubernetes uses it for API declarations and validation policies.

Proxmox Virtual Environment — The open-source hypervisor platform that Nick Steinmetz manages entirely through Swamp in his homelab, including Discord-driven VM creation.

terraform-aws-modules/vpc — Anton Babenko's widely-used Terraform VPC module, referenced by Paul as an example of human-centric IaC design with 237+ inputs that agents struggle to navigate.

We are always happy to answer any questions, hear suggestions for new episodes, or hear from you, our listeners.

DevSecOps Talks podcast LinkedIn page

DevSecOps Talks podcast website

DevSecOps Talks podcast YouTube channel

We are always happy to answer any questions, hear suggestions for new episodes, or hear from you, our listeners.

DevSecOps Talks podcast LinkedIn page

DevSecOps Talks podcast website

DevSecOps Talks podcast YouTube channel

We are always happy to answer any questions, hear suggestions for new episodes, or hear from you, our listeners.

DevSecOps Talks podcast LinkedIn page

DevSecOps Talks podcast website

DevSecOps Talks podcast YouTube channel

We are always happy to answer any questions, hear suggestions for new episodes, or hear from you, our listeners.

DevSecOps Talks podcast LinkedIn page

DevSecOps Talks podcast website

DevSecOps Talks podcast YouTube channel

We are always happy to answer any questions, hear suggestions for new episodes, or hear from you, our listeners.

DevSecOps Talks podcast LinkedIn page

DevSecOps Talks podcast website

DevSecOps Talks podcast YouTube channel

We are always happy to answer any questions, hear suggestions for new episodes, or hear from you, our listeners.

DevSecOps Talks podcast LinkedIn page

DevSecOps Talks podcast website

DevSecOps Talks podcast YouTube channel

We are always happy to answer any questions, hear suggestions for new episodes, or hear from you, our listeners.

DevSecOps Talks podcast LinkedIn page

DevSecOps Talks podcast website

DevSecOps Talks podcast YouTube channel