Learn AI in Bits

A model with no company name attached to it appeared free on OpenRouter on August 20, 2026, and an early coding test made it look like a serious challenger to Claude and GPT. This episode covers Ox Alpha, the anonymous stealth model developers spent the week testing: what OpenRouter and independent researchers have confirmed about it, why a larger benchmark run changed the headline number, and what an anonymous provider means for anyone thinking about sending it sensitive work.

Ox Alpha showed up on OpenRouter under the placeholder ID stealth/ox-alpha, also reachable through coding tools like OpenCode and Cline, with OpenRouter naming the provider only as an anonymous third party. It ships with a context window of 1,048,576 tokens, a maximum output of 131,072 tokens, and support for text, image, and video input, built for coding, sustained agentic work, and production use with tool calling and structured output. The provider hasn't disclosed a parameter count for the underlying model.

The coding benchmark is where the episode gets specific. Developer Ben Davis first ran Ox Alpha through a 10-task DeepSWE coding test and reported an 80% pass rate, ahead of roughly 65% for Claude and 52% for GPT-5.6 in that same informal comparison. A larger follow-up test brought the score down to about 63% — a small enough sample swing to make clear that a ten-task result isn't a settled ranking, and that early numbers from an unverified model deserve a second test before they get repeated as fact.

Identity is the other open question. Researchers comparing Ox Alpha's tokenizer behavior, video handling, and response fingerprints have pointed toward Zhipu AI's GLM family as a possible origin, but no company has confirmed building it, and the provider has stayed anonymous through the entire preview period. The episode also covers the privacy angle: OpenRouter states that prompts and completions sent to Ox Alpha are retained by the provider, even though they aren't used for training, which is worth factoring in before sending confidential or proprietary material to a model nobody has claimed.

Sources & References

OX Alpha: The Anonymous Frontier Model - Comprehensive Technical Analysis, Local AI Zone — https://local-ai-zone.github.io/blog/ox-alpha-stealth-model-comprehensive-analysis.html

Ox Alpha likely GLM-5.3: the mystery model's fingerprint, OrcaRouter — https://www.orcarouter.ai/blog/ox-alpha-stealth-model-what-we-know

Ox Alpha: Anonymous 1M-Context Model Hits No. 2 on OpenCode in Three Days, The Cherry Creek News — https://thecherrycreeknews.com/ox-alpha-stealth-model-openrouter-benchmarks-analysis-cherry_creek/

Ox Alpha - API Pricing & Providers, OpenRouter — https://openrouter.ai/stealth/ox-alpha

Ox Alpha on OpenRouter: Free 1M Stealth Model (Aug 2026), explainx.ai — https://explainx.ai/blog/openrouter-ox-alpha-stealth-model-august-2026

Voice narration is AI-generated.

Anthropic launched Claude Academy on March 2, 2026, a free, self-paced learning platform at academy.claude.com. This episode covers what's actually in it, how the courses are organized, and a closer look at the "AI Fluency" framework Anthropic puts at the center of the whole thing, including a worked walkthrough of its four core competencies applied to a real task.

Claude Academy started with 13 courses and has grown to around 20, ranging from introductory material for people who've never used Claude to technical deep dives on the Claude API and Claude Code, plus dedicated pathways for K-12 educators and nonprofits. The catalog is organized by how you'd actually use Claude rather than by difficulty: tracks cover Claude.ai, Claude Cowork, the developer-facing Claude Code and Claude Platform tools, and Claude Tag inside Slack. On the technical end, "Building with the Claude API" runs 84 lectures across 8 hours with 10 quizzes, with matching courses for running Claude through Amazon Bedrock and Google Vertex AI. There's also a dedicated course on the Model Context Protocol and one on model capabilities and limitations. Every course awards a certificate on completion, and none of it requires a credit card.

The course Anthropic points newcomers to first isn't technical at all. "AI Fluency: Framework and Foundations" teaches what Anthropic calls the four D's: Delegation, deciding whether, when, and how to engage AI on a task; Description, communicating a goal clearly enough to get a useful result; Discernment, accurately judging whether an AI's output is actually good; and Diligence, taking responsibility for what gets done with AI and how. The episode walks through all four applied to drafting a difficult client email, and touches on the framework's three modes of human-AI interaction: automation, augmentation, and agency.

The larger point is that the four D's aren't specific to Claude or to any one platform. They're a vocabulary for a kind of judgment that doesn't show up in a changelog, and having names for each step gives a person, or a team, a way to talk about exactly where a task with AI went right or wrong.

Sources & References

Claude Academy — https://academy.claude.com/

Claude course catalog — https://claude.com/resources/courses

Claude tutorials — https://claude.com/resources/tutorials

Anthropic, AI Fluency Framework and Foundations PDF — https://www-cdn.anthropic.com/b383cf6baddbfc72fdf8b0ed533a518e2872d531.pdf

Cryptobriefing, Anthropic launches Claude Academy to enhance AI learning resources — https://cryptobriefing.com/anthropic-launches-claude-academy/

Voice narration is AI-generated.

Ask a developer to describe their AI stack and most people picture one model wired into one clever autonomous agent. Production data shows something messier. This episode maps what the AI development stack actually looks like in 2026 — models, databases, languages, and the connective tissue holding it together — using current developer telemetry from Datadog and a joint 2026 industry survey instead of marketing claims.

On models, Datadog's analysis of AI usage across more than a thousand customers found OpenAI used by 75% of organizations a year ago, down to 63% now, while Anthropic's Claude gained 23 percentage points and Google's Gemini gained 20 over the same stretch. More than 70% of organizations now run three or more models side by side, picking each one for quality, speed, cost, or the task at hand rather than staying locked into a single provider. Frameworks like LangChain sit between those models and the rest of an application, but adoption is still far from universal: Datadog found agent-framework usage at only about 18% of organizations in early 2026. Production telemetry also shows most of what gets labeled an "agent" is doing far less than the word implies: 59% of agentic applications made just a single service call, and only 18% chained three or more together.

On the database layer, Datadog found about 60% of organizations run at least one production Postgres instance, and its vector-search extension, pgvector, was the fastest-growing non-bundled Postgres extension in their data, up 24% between December 2025 and May 2026 — but still early, with fewer than 6% of organizations and under 1% of tables actually using a vector index, leaving dedicated vector databases like Pinecone, Weaviate, and Qdrant to handle retrieval at real scale.

On languages, Python remains the leading language ecosystem for accessing Postgres in production, up 33% over two years, while Node and TypeScript stay important for web applications, APIs, and streaming interfaces — there's no single required split, and most teams use each language where it naturally fits.

The episode closes on the economics, drawn from a joint 2026 AI engineering survey by Notion, Amplify Partners, and Vercel: about three-quarters of respondents said cost regularly or sometimes changes how they use AI, inference and model serving is the most commonly purchased layer rather than built in-house, and only 28% of eligible calls actually use prompt caching despite the cost savings on offer. The same survey found 97% of respondents reporting positive effects from AI, alongside 93% reporting at least one negative effect — skill erosion the most common at 29%, followed by higher review burden at 28% and brittle, incident-prone code at 22%.

Sources & References

Datadog, State of AI Engineering — https://www.datadoghq.com/state-of-ai-engineering/

Datadog, State of Postgres — https://www.datadoghq.com/state-of-postgres/

2026 AI Engineering Survey, Notion, Amplify Partners, and Vercel — https://www.notion.com/lp/ai-engineering-survey

Amplify Partners, The 2026 AI Engineering Report — https://www.amplifypartners.com/blog-posts/the-2026-ai-engineering-report

Voice narration is AI-generated.

One click on a link, and Microsoft's own Copilot could read a user's email, Google Drive, and calendar, then quietly send it to a stranger. This episode explains prompt injection, the security flaw unique to AI agents, through a real Microsoft Copilot vulnerability that was patched after sitting open for months.

A model reading text can't reliably tell the difference between an instruction from its actual user and an instruction hidden inside content it's processing, a webpage, an email, a shared document. If that content contains hidden text telling the model to ignore its real instructions and leak data instead, a model with no built-in sense of who's allowed to give it orders can follow that instead. Researcher Simon Willison named the exact combination that makes this dangerous the "lethal trifecta": an agent with access to private data, exposure to untrusted content, and the ability to communicate externally. Microsoft's Copilot had all three.

Security researchers at Varonis Threat Labs found a flaw they named CoSnitch: a link combining two URL parameters, one that pre-filled Copilot's chat box with a hidden prompt and one, called autorun, that executed it automatically the moment the page loaded, with no typing required and no way to stop it by closing the tab. Once running, the injected prompt used Copilot's own built-in web-fetching tool to pull email subject lines and bodies, Google Drive filenames, calendar details, chat history, and Copilot's own saved memory, then send it out encoded to dodge detection filters. A companion flaw let attackers plant instructions directly into that saved memory, persisting through a password change, a revoked session, or even a new device.

Varonis reported the issue to Microsoft on December 31, 2025. Microsoft shipped a partial fix that February and didn't close the vulnerability completely until August 18, 2026, about seven and a half months later, on a bug Microsoft itself rated critical. The episode explains why that timeline matters: prompt injection doesn't patch the way an ordinary bug does, because the underlying problem, a model's inability to fully separate its user's instructions from anyone else's, isn't unique to Copilot. Any AI assistant wired into private data, exposed to the open web, and able to send data back out carries the same three ingredients, regardless of which company built it.

Sources & References

The Hacker News, Microsoft Copilot Personal Flaws Could Let One Click Exfiltrate Data From Connected Apps — https://thehackernews.com/2026/08/microsoft-copilot-personal-flaws-could.html

Computerworld, Microsoft finally patches critical one-click Copilot vulnerability, almost eight months after learning of it — https://www.computerworld.com/article/4211325/microsoft-finally-patches-critical-one-click-copilot-vulnerability-more-than-eight-months-after-learning-of-it.html

Simon Willison, The lethal trifecta for AI agents: private data, untrusted content, and external communication — https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/

Voice narration is AI-generated.

Paste a 200-page PDF into a chatbot and it answers questions about page 180 within seconds. That feels instant and free. Neither is quite true, and this episode explains what a context window actually is, what happens inside a model when it processes a huge document, and why a bigger window always comes with a cost.

A context window is the total amount of text, measured in tokens, a model can hold at once across the prompt, any pasted document, and its own reply. Flagship models from OpenAI, Google, and Anthropic have all converged around a roughly one-million-token window in 2026, up from the 4,000- and 8,000-token limits those same companies shipped just a few years earlier. Inside that window, a transformer compares every token against every other token to decide what to attend to, so doubling the amount of text roughly quadruples the compute required, which is why OpenAI's own pricing sheet charges more once a request crosses into what it calls long-context territory.

Size brings a second cost that has nothing to do with money: research on "lost in the middle" behavior shows models are measurably worse at using information buried in the center of a long context than information at the start or end, and newer testing on today's huge-context models found that gap gets worse, not better, as windows grow, with 10 to 25 percent accuracy loss on facts placed mid-document.

The episode works through the actual math on a 200-page PDF: at roughly 500 words a page and OpenAI's own conversion rate of about 0.75 words per token, that document comes out to roughly 130,000 tokens and around 33 cents in input costs before the model writes a single word back. It also covers a concrete real-world response to that tradeoff: OpenAI's own coding tool, Codex, deliberately caps how much of GPT-5.6's advertised 1.05-million-token window it will actually use at once, capping it at 272,000 tokens specifically because of mid-document accuracy loss and runaway cost. The practical takeaway is simple: put what matters most at the start or end of a long prompt, not buried in the middle.

Sources & References

OpenAI, Key concepts (token-to-word conversion) — https://developers.openai.com/api/docs/concepts

OpenAI, API pricing — https://developers.openai.com/api/docs/pricing

Codex Knowledge Base, The Context Window Gap: Why Codex CLI Caps GPT-5.6's Million-Token Window at 272K — https://codex.danielvaughan.com/2026/07/20/context-window-gap-codex-cli-gpt56-advertised-vs-effective-budget-compaction-strategy/

hidekazu-konishi.com, LLM Context Window Growth Timeline — https://hidekazu-konishi.com/entry/llm_context_window_growth_timeline.html

Voice narration is AI-generated.

Meta and Mistral give away state-of-the-art AI models for free, no purchase or subscription required. That sounds like open source software, but it isn't quite the same thing, and the difference has real consequences for anyone building on top of these models.

This episode explains what "open weight" actually means and how it differs from open source AI. A model's weights are the millions of parameters it learned during training; publishing them lets anyone download and run the finished model. The Open Source Initiative, the group that defines the term for software, says that alone doesn't qualify as open source: a model only earns that label if it also comes with the training code and real details about the training data, the sources that shaped it, not just the finished result.

The distinction gets concrete in the actual license text. Meta's Llama 4 Community License Agreement requires any company with more than 700 million monthly active users in the preceding month to request a separate license from Meta, and it blocks using Llama's own outputs to train a competing model, restrictions a genuine open source license like Apache or the ones covering Linux would never include. Mistral, the French AI company, ships most of its core models under Apache 2.0, a real open source license with no usage cap, while other open-weight releases, like Google's Gemma, carry their own restrictions, including the right to remotely cut off access if usage rules are broken.

The episode also covers why these companies give away models that cost hundreds of millions of dollars to train. For Meta, a free model functions as a large-scale recruiting tool, builds a developer ecosystem, and pressures closed competitors like OpenAI and Anthropic on price, all while Meta's core revenue stays tied to advertising rather than model access. Mistral co-founder Guillaume Lample frames the company's reasoning around independence: "We don't want AI to be controlled by only a couple of big labs," a statement that also reads as a pitch for European AI sovereignty against dominant American labs.

Sources & References

Open Source Initiative, Open Weights: not quite what you've been told — https://opensource.org/ai/open-weights

Meta, Llama 4 Community License Agreement — https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama4/LICENSE

TechCrunch, Mistral closes in on big AI rivals with Mistral 3, open-weight frontier and small models — https://techcrunch.com/2025/12/02/mistral-closes-in-on-big-ai-rivals-with-mistral-3-open-weight-frontier-and-small-models

Voice narration is AI-generated.

Why does ChatGPT answer some questions instantly and pause to "think" on others? Most people assume they're always talking to the same AI brain. In reality, a router decides in a fraction of a second which model actually handles your message, and that decision shapes both the quality of the answer and what it costs to produce.

This episode explains model routing, the invisible layer sitting in front of most modern AI products. OpenAI built a router directly into GPT-5 when it launched in 2025, and it's still the default behavior in ChatGPT: the system looks at your conversation type, how complex the request appears, whether tools are needed, and explicit signals like typing "think hard about this," then sends easy requests to a fast, efficient model and harder ones to a deeper reasoning model. The router keeps improving over time based on which answers users actually preferred and how often each model got things right.

Routing isn't limited to one company's own lineup. OpenRouter, a platform that sits in front of roughly 400 models from OpenAI, Anthropic, Google, and dozens of other providers, runs its own "auto" router that classifies a prompt by task type and picks a model based on real usage patterns from its community over the trailing week. OpenRouter said it was serving about 8 million users by this spring, and on August 19, 2026, days before this episode, Stripe confirmed it's buying the company in a deal reported at more than $7 billion, a bet that routing between AI models is becoming as central to the internet as routing payments already is.

The episode works through a real cost comparison to show what's actually at stake in a routing decision: on OpenAI's current pricing, the efficient GPT-5.4 nano model runs about twenty cents per million input tokens, while the full GPT-5.4 model runs two dollars and fifty cents for the same volume, more than twelve times as much. That gap is the economic reason routing exists, and it explains why most routed products still let a user override the router's guess when the stakes are high enough to want a specific model by hand.

Sources & References

OpenAI, Introducing GPT-5 — https://openai.com/index/introducing-gpt-5/

OpenRouter, How Model Routing Works — https://openrouter.ai/blog/insights/model-routing/

Morph, OpenAI API Pricing — https://www.morphllm.com/openai-api-pricing

TechCrunch, Stripe will reportedly acquire AI gateway startup OpenRouter for $7B+ — https://techcrunch.com/2026/08/16/stripe-will-reportedly-acquire-ai-gateway-startup-openrouter-for-7b/

Axios, Stripe confirms OpenRouter acquisition — https://www.axios.com/pro/fintech-deals/2026/08/19/stripe-openrouter-acquisition

Voice narration is AI-generated.

AI agent systems are moving beyond a single model repeatedly calling tools. Two useful concepts for understanding that evolution are loop engineering and graph engineering.

Loop engineering focuses on the behavior of an individual agent over time. An agent receives context, chooses an action or tool call, observes the result, and decides what to do next. The engineering challenge is making that cycle reliable: defining progress, providing useful feedback, handling failures, managing context, setting budgets, and deciding when the agent should stop. OpenAI's Codex documentation describes the agent loop as the core logic that orchestrates the user, model, and tools. OpenAI also notes that a single turn can contain many model-and-tool iterations and that context-window management becomes an important responsibility.

Graph engineering addresses a different problem: coordinating a larger system of steps, agents, tools, branches, dependencies, and shared state. Instead of leaving every next action to one agent, the system can represent the workflow explicitly as a graph. Nodes can represent tasks or specialized agents, while edges describe dependencies or transitions.

For example, a research system might have separate research, data-analysis, writing, and review nodes. Research and data analysis could run in parallel, their results could feed a writing node, and the draft could move to a review node. If the reviewer finds a problem, the graph can route the work back for revision.

The important distinction is that loop engineering is primarily about iteration and behavior, while graph engineering is about structure and coordination. They are complementary. A node inside a graph can contain its own agent loop. The graph determines when that node runs, what it receives, and where its output goes; the loop determines how the agent works through its assigned task.

The term "graph engineering" is still emerging rather than a universally standardized discipline. Many underlying techniques have existed for years in workflow orchestration, state machines, directed graphs, DAGs, and multi-agent systems. What is new is applying the graph-engineering framing directly to increasingly complex AI agent systems.

A paper published August 21, 2026, "Graph Engineering in the Era of LLM Agents: From Individual Intelligence to System Intelligence," explicitly proposes graph engineering as an emerging paradigm for organizing complex agent systems. The paper argues that some tasks require heterogeneous expertise, parallel execution, independent verification, and persistent state that can exceed the organizational capacity of a single agent.

REFERENCES

OpenAI — Unrolling the Codex agent loop

https://openai.com/index/unrolling-the-codex-agent-loop/

OpenAI — The next evolution of the Agents SDK

https://openai.com/index/the-next-evolution-of-the-agents-sdk/

OpenAI — Harness engineering: leveraging Codex in an agent-first world

https://openai.com/index/harness-engineering/

OpenAI — An open-source spec for Codex orchestration: Symphony

https://openai.com/index/open-source-codex-orchestration-symphony/

Anthropic — Harness design for long-running application development

https://www.anthropic.com/engineering/harness-design-long-running-apps

Feng et al. — Graph Engineering in the Era of LLM Agents: From Individual Intelligence to System Intelligence

https://arxiv.org/abs/2608.21156

Macedo — What makes prompts a graph: necessary and sufficient conditions for prompt graph engineering

https://arxiv.org/abs/2607.27578

Voice narration is AI-generated.

For about fifteen years, computer science was one of the safest majors to recommend. The 2026 data checks that advice: enrollment, unemployment, hiring trends, and a real curriculum redesign, for parents, students, current CS majors, and professors alike.

The National Student Clearinghouse reported computer and information sciences enrollment fell 8.4% at four-year schools and 11.2% at two-year colleges in spring 2026, with graduate enrollment down 14%, even as total enrollment grew nationally. A Computing Research Association survey of 134 academic leaders found 62% of programs saw undergraduate enrollment drop, typically by 11% to 15%. Computer science now has the lowest persistence rate of the ten most popular bachelor's degrees, at 85%, versus 93% for engineering.

That same CRA survey ties the decline to fear that AI will eliminate software engineering jobs, and two thirds of computing departments say recent grads are struggling to find work. Federal Reserve Bank of New York data, drawn from Census figures on graduates in their early twenties, puts computer science unemployment at 7% and computer engineering at 7.8%, fifth and second highest of 73 majors tracked. Entry level hiring has tightened: the junior share of tech employment fell from 15% to 7% in three years, and Microsoft and Google both say roughly 30% of their code is now AI generated, with Meta targeting 50%.

But the data also cuts the other way. Computer science graduates who do find work are underemployed at under 20%, versus roughly 42% for graduates overall, and the major still ranks third among the most in demand bachelor's degrees. IBM says it's tripling entry level technical hiring. Some of the pullback likely traces to pandemic era over hiring correcting itself rather than AI, and the enrollment drop looks more like a shift toward specialization: cybersecurity, data science, computer engineering, and standalone AI programs are all growing inside the same departments.

The University of Iowa redesigned its computer science curriculum for fall 2026 around 2023 to 2024 Association for Computing Machinery guidelines, adding required ethics and intro to AI courses and combining separate statistics, probability, and linear algebra classes into one applied math course. That redesign lines up with what hiring managers say they actually want: shipped, deployed projects over GPA, real maintenance experience, and fluency working alongside AI tools instead of trusting or ignoring them. Direct guidance follows for each audience, plus cybersecurity, data science, computer engineering, and physical infrastructure fields as the alternatives current data points toward.

Sources & References

National Student Clearinghouse, computer science enrollment data: https://www.studentclearinghouse.org/nscblog/computer-science-enrollment-is-cooling/

CRA CERP Pulse Survey: https://cra.org/crn/2025/10/cerp-pulse-survey-a-snapshot-of-2025-undergraduate-computing-enrollment-patterns/

NY Fed, College Labor Market data: https://www.newyorkfed.org/research/college-labor-market

VnExpress, U.S. computer science graduates unemployment rate: https://e.vnexpress.net/news/tech/tech-news/us-s-computer-science-grads-face-5th-highest-unemployment-rate-5048578.html

Built In, computer science degree decline and AI: https://builtin.com/articles/computer-science-degree-decline-ai

Extern, computer science job market guide 2026: https://www.extern.com/post/computer-science-job-market-2026-guide

Boise State University, AI and computer science demand: https://www.boisestate.edu/coen-cs/2026/03/10/more-ai-means-more-computer-science/

Ardura Consulting, junior developer hiring crisis: https://ardura.consulting/blog/junior-developer-crisis-2026-why-companies-stopped-hiring-entry-level/

U. of Iowa, new CS curriculum for fall 2026: https://cs.uiowa.edu/news/2026/06/new-computer-science-curriculum-fall-2026

Voice narration is AI-generated.

Can a single tech-savvy person actually build and maintain an AI agent for a small business, or is that just a demo-video fantasy? A full, concrete walkthrough answers that: a fictional flower shop called Rosewood Flowers automates its same-day delivery scheduling using Claude, built from an agent loop, a custom skill, and four real-world connectors, then stress-tested against the costs, maintenance, and failure modes that never make it into a product demo.

In plain terms: the difference between a tool (a single action Claude can take), a connector built on the Model Context Protocol, or MCP (how Claude reaches external systems like Shopify, Google Calendar, and a mapping service), a skill (a packaged, progressively-loaded set of business rules Claude only pulls into context when relevant, built using Claude's Agent Skills framework and SKILL.md files), and an agent loop using subagents from Claude's Agent SDK, where a main agent delegates focused tasks, intake, routing, and customer notification, to specialized subagents with their own tools and instructions.

Listeners get a full order walked end to end through the system, plus a harder edge case that deliberately escalates to a human instead of guessing. It also gets specific about where this kind of system actually runs: building it requires Claude API access through Anthropic's developer console, a separate product and separate billing from a Claude.ai subscription like Pro or Max, and the skill and orchestration code live as ordinary files, a SKILL.md file and Python or TypeScript scripts, in a code repository like GitHub. Two hosting paths are covered: self-hosting the code on a small server or container, or using Managed Agents, the hosted API Anthropic opened to developers in a 2026 public beta that runs the agent and sandbox directly with no server to maintain.

The math on ongoing token costs uses Claude's actual current Sonnet pricing, two dollars per million input tokens and ten dollars per million output tokens, and flags that this is an introductory rate expiring September 1, 2026, a concrete example of how AI infrastructure costs shift over time. It also covers realistic build time (two to four weeks part-time), the skills a builder actually needs, and the maintenance nobody advertises: expiring OAuth tokens, connector APIs changing, model version updates requiring re-testing, and business rule changes that require editing the skill file before the next busy weekend arrives.

This is a technical explainer built for listeners who want to understand agents, skills, connectors, and hosting through one fully worked, honest example rather than marketing language.

Sources & References

Claude Agent Skills overview (SKILL.md, progressive disclosure) — https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview

Claude MCP connector documentation — https://platform.claude.com/docs/en/agents-and-tools/mcp-connector

Claude Agent SDK subagents documentation — https://code.claude.com/docs/en/agent-sdk/subagents

Hosting the Agent SDK (self-hosted deployment options) — https://code.claude.com/docs/en/agent-sdk/hosting

Claude.ai subscription vs. API/Console access — https://support.claude.com/en/articles/9876003-i-have-a-paid-claude-subscription-pro-max-team-or-enterprise-plans-why-do-i-have-to-pay-separately-to-use-the-claude-api-and-console

Anthropic Managed Agents public beta overview — https://www.roborhythms.com/anthropic-managed-agents-2026/

Claude API pricing overview, 2026 — https://www.cloudzero.com/blog/claude-pricing/

Voice narration is AI-generated.