Act of Intelligence

Note: this episode was recorded before the release of Opus 5.5 and Sol 6.1. Ajay and Andrew's model preferences have changed since then.

Andrew spent three hours answering a Stanford-derived question bank so a digital clone of himself could sit at the top of his agent fleet, and Ajay wants to know what that clone actually is. Is it a twin, a clone, or a branch of you frozen on one particular day? The two of them get pedantic about the words, philosophical about whether any questionnaire captures a person who makes decisions on bad sleep and pizza, and then very practical: the clone answers blockers with Andrew's values, escalates only the irreversible one-way-door decisions, and has cut the human interruptions from roughly twenty a day to one. Along the way: four days of agents writing tests for the tests, dropping every agent to low effort to ship an MVP, a Minecraft server project reaching its first testable milestone, a bare-metal Kubernetes cluster instead of AWS, which model belongs in which seat, and the recurring question of whether any of this is intelligence or just a very good act.

In this episode:

Why a digital clone: removing the judgment calls that are obvious to a human but invisible to an agent

The Stanford digital-twin study, its question bank, and how the interview was extended for Andrew's projects

Four days of CI churn, tests for the tests, and the decision to set every agent to low effort

Twin versus clone: the moment the copy is made, the experiences diverge

Whether a three-hour interview can capture a person whose judgment shifts with sleep, food, and mood

One-way door decisions: recoverable mistakes are lessons, irreversible ones get escalated to the human

Logging every session, what to keep raw, what to summarize, and why the judgment log is the real asset

Inside Reign, Andrew's agent harness: an owner tier, delegated thrones, a council of sub-agents, and a warden

Interruptions down from about twenty a day to one, and ten of sixteen features shipped in a day

A Minecraft server project reaches MVP and pauses on a one-way door: the server architecture

Talos Linux, Kubernetes, and a single bare-metal node at OVH, the provider Andrew is weighing, growing into a three-node cluster

How to start your own clone: a Stanford question bank, roughly 150 questions, and an Andrew.md that filters the model

Standing rules for the clone: no decisions affecting friends or family, and ignore anything marked urgent

Fable for the human-facing work, Codex for the code, and why Opus 5 can be hard to read

Is it acting or intelligent? This week: a double act

Chapters:

(00:00:00) - Teaser: four days, zero features

(00:00:28) - Cold open: a digital twin joins the agent fleet

(00:03:25) - Building a compass so agents can clear blockers

(00:05:47) - Logging every decision for the long view

(00:08:40) - Starting from the Stanford digital-twin question bank

(00:10:17) - Four days of CI churn, then effort set to low

(00:14:19) - Twin or clone? Getting pedantic

(00:16:17) - Can three hours of questions capture your essence?

(00:22:47) - One-way doors: escalate the irreversible, forgive the rest

(00:28:31) - Your judgment log is the real asset

(00:35:45) - Inside Reign, the harness: thrones, a council, and a warden

(00:40:08) - From twenty pings a day down to one

(00:43:44) - Rigor restarts: ten of sixteen features in a day

(00:46:17) - Luxwander hits MVP and waits on server architecture

(00:47:38) - Talos, Kubernetes, and OVH bare metal instead of AWS

(00:55:51) - "I am the bottleneck for all my projects"

(00:58:13) - How to build your own digital clone

(01:02:17) - The partner test and a no-family-decisions rule

(01:04:19) - Why the clone ignores anything marked urgent

(01:05:16) - Which model for which seat

(01:07:02) - Opus verbosity, quantized replies, and Codex for the code

(01:10:10) - Bringing on guests who know more

(01:10:35) - Acting or intelligent? A double act

(01:11:47) - Twenty years of epic technology

What does it take to stop babysitting your coding agents? Ajay and Andrew open with how big-tech office politics quietly manage people out, then get into the real news: Luxwander finally has a true progress number (95%), and the workflow that made it measurable has been split out into Rigr — a spec-first system with assurance tiers that let you dial expected quality from NASA-grade down to digital handyman. Along the way: a Minecraft server with a physical gold economy, why front-loading specifications kills a hundred rounds of code review, parallel CI as the new agentic bottleneck, the multi-repo versus monorepo argument, and a closing debate on whether today's models are intelligent or just acting.

In this episode:

How people get managed out, and what an IC can move without authority

Luxwander at 95 percent, and what a real progress bar took

Minecraft Nations: guilds, territory, and a physical in-game gold economy

Assurance tiers: matching AI effort and cost to the stakes of the job

Rigr: specs, slices, and contracts before a line of code is written

Parallel CI, multi-repo orchestration, and the monorepo debate

Are we the intelligent ones, or is the AI acting?

Chapters:

(00:00:00) - Teaser: launch a rocket on tier five

(00:00:18) - Office politics and how people get managed out

(00:06:37) - What an IC can actually move without authority

(00:07:42) - Luxwander at 95%: finally a real progress number

(00:10:36) - From "just build it" to eighteen modules

(00:11:52) - Minecraft Nations: guilds, territory, a physical gold economy

(00:19:28) - Assurance tiers: NASA at one, digital handyman at five

(00:25:39) - Why Rigr got split out of Luxwander

(00:31:39) - Zooming out to an orchestration layer over the agents

(00:38:30) - Front-loading specs to kill a hundred rounds of review

(00:40:55) - Modules, slices, contracts, and isolation

(00:51:16) - Parallel CI becomes the agentic bottleneck

(01:00:25) - Multi-repo orchestration and the monorepo argument

(01:06:23) - Opinionated defaults, and why design stays human

(01:08:43) - Tier thought experiments: rockets and cancer screening

(01:10:58) - Are we the intelligent ones?

Ajay and Andrew start with a cat-induced re-recording and end somewhere they did not expect: the first real "yes" to the show's core question. In between they dig into the software engineering loop, why contracts and modularization make AI-assisted work parallelizable, how to tell when a system actually needs to be split up, and why observability and flexibility are the two tenets they will not give up. They compare Codex, Claude, Fable, and GLM in real use, from usage limits and precision to one-shotting and transparent peak-hour pricing, trade notes on adopting the GSD workflow without losing their own, and watch Sonnet 5 debug an AWS permissions rabbit hole on its own.

In this episode:

Contracts and modularization, and why they make AI-assisted work parallelizable

How to tell when a system actually needs to be split up

Codex vs. Claude vs. Fable vs. GLM in real, day-to-day use

Why observability and flexibility are the two tenets of the loop

Adopting a shared workflow without giving up a hand-built one

The first real "yes" to the show's question: act, or intelligence?

Chapters:

(00:00:00) - Welcome (and a cat-induced take two)

(00:02:42) - The "software engineering loop": planning and parallelization

(00:05:20) - What's a contract? APIs, JSON, and Amazon Bedrock

(00:08:04) - Versioning contracts: 1.0 and breaking changes

(00:12:09) - Enforcing contracts: validation, CI, and TDD

(00:14:39) - Don't modularize too early: velocity as the signal

(00:17:23) - Model limits and generosity (and Grok 4.5 lands)

(00:20:24) - Codex's precision vs. Claude's fill-in-the-blanks

(00:22:39) - Concise vs. detailed, and OpenAI's Sol/Luna/Tara

(00:25:30) - Trying GLM from z.ai, and transparent peak-hour pricing

(00:32:09) - A bash wrapper to swap models into the Claude harness

(00:34:47) - Back to Claude: why Fable wins on efficiency

(00:37:07) - Is Fable one-shotting because SWE principles got trained in?

(00:38:19) - Sonnet 5 down an AWS SAM permissions rabbit hole

(00:42:41) - Observability for coding agents (LangFuse, Braintrust)

(00:45:16) - Adopting GSD: hooks, overlap, and what to steal

(00:53:03) - Two tenets of the loop: observability and flexibility

(00:54:38) - Blind spots, rituals, and what GSD surfaced

(01:01:36) - Act or intelligence? The first "yes"

Ajay and Andrew get pulled off their notes and into a bigger question: as coding assistants get better at running themselves, what is left for us to do? Fresh off a week of hands-on time with Claude Code and Codex, they compare the limits, the harnesses, and the agents that spawn their own agents, then land somewhere surprisingly practical: the age of the "digital handyman," where anyone can build bespoke software for their own life. Along the way they get into the LLM wiki as a second brain, why Anthropic may be out-earning OpenAI on a fraction of the users, whether vibe coding is dead or just getting started, and the closing verdict that the best handyman might be hands off.

In this episode:

The LLM wiki as a second brain that lives beyond the model's context window

Why the harness, not the model, is becoming the real product

Anthropic vs. OpenAI, and what "winning on revenue" actually means

Guardrails, newbies, and whether the software development lifecycle still matters at home

Building your own finance app when Monarch and Plaid keep breaking

Codex's "goal" that runs for days, and why the best handyman might be hands off

Chapters:

(00:00:00) - Welcome: what "Act of Intelligence" is about

(00:02:08) - The LLM wiki: a second brain beyond the context window

(00:04:36) - How the wiki works: concepts, synthesis, and "runoff"

(00:16:35) - A home-OS detour, and modularizing for the human mind

(00:26:28) - Claude Code's limits vs. Codex's generosity

(00:27:31) - Anthropic vs. OpenAI: revenue on a fraction of the users

(00:32:49) - Wrappers over models: the harness is the value

(00:34:13) - The desktop app that spawned and reviewed its own agents

(00:38:32) - Guardrails, newbies, and "is vibe coding dead?"

(00:44:24) - When software isn't a moat (and a billion habit trackers)

(00:46:57) - Monarch, Plaid, and rolling your own finance app

(00:49:30) - Pillars of Wealth and the home finance assistant

(00:52:19) - The digital age of the handyman

(00:56:51) - Now you're the product manager (and trimming the tools)

(01:03:45) - Codex's "goal": the run that wouldn't stop

(01:08:29) - "The best handyman is hands off"

(01:11:07) - Opinionated harnesses, and the future they're building

(01:14:11) - Mirage or oasis? The close

If coding is "solved," is software engineering? Ajay Medury (software engineer) and Andrew Sierota (systems engineer) pick up where Episode 1 left off and get into the part that isn't solved: judgment. They trade notes on why weekly usage limits have quietly become the real project budget, what it's like to build a sharded Minecraft world solo as both product manager and principal engineer, what Amazon's New World got wrong about scale, running decorrelated multi-model code reviews, and what an AI "skill" actually is. It might all just be an act of intelligence.

In this episode:

• Why "coding is solved" but software engineering isn't, and why judgment is the expensive part
• The new bottleneck: weekly subscription usage limits as a hard budget
• Breaking a big build into modules and submodules, and shrinking scope to actually ship
• Wearing every hat at once: product manager and principal engineer
• New World and the problem of scale at launch
• Large codebases, heavy test coverage, and review rounds that exposed process gaps
• Pre-flight checks, linters, and spec-tracing that cut review loops down
• Decorrelated reviews: several different models reviewing blind, then taking the union of findings
• What an AI "skill" is: system prompts, user prompts, and guardrails for long workflows
• Severity tiers for findings: blockers, warnings, defers, suggestions, and nits
• Why systems admins, as generalists by trade, may be an ideal audience for these tools

Chapters:

• (00:00:00) - Real intelligence, or just an act?
• (00:01:02) - Coding is solved; software engineering isn't
• (00:02:02) - Keeping up with the release pace
• (00:05:27) - Vibe coding vs. a repeatable process
• (00:06:40) - Usage limits are the new budget
• (00:08:40) - Breaking the build into modules
• (00:12:18) - A sharded world, every hat on one builder
• (00:17:12) - New World and the problem of scale
• (00:25:04) - 200k lines and a 90-round review
• (00:27:20) - Pre-flight checks that cut 90 rounds to 5
• (00:30:49) - The podcast's own local-GPU pipeline
• (00:33:27) - Learning by asking "what do you mean?"
• (00:35:35) - When a large agent run burned through the budget
• (00:38:05) - What is a skill, really?
• (00:48:05) - Skills as guardrails for long workflows
• (00:51:09) - Severity tiers: blockers to nits
• (00:54:34) - Why sysadmins are ideal builders
• (00:56:29) - A long-running Minecraft community, the real driver
• (01:00:56) - Closing: an act of intelligence

Welcome to Act of Intelligence — a new podcast about AI tools, workflows, and the philosophy of building with them, hosted by Ajay Medury and Andrew Sierota. In this pilot, a software engineer and systems engineer compare notes on what it's actually like to build real software with AI coding agents like Claude Code.

They get into plan mode and context management, the explosion of open-source "memory" tools for Claude, and the build-vs-buy trap of reinventing the wheel. Along the way: why "writing code is a solved problem" but shipping working software isn't, how the job is shifting from coder to technical PM / systems architect, the engineering fundamentals (DRY, modular design, conventions, spec-driven development) that suddenly matter more in the AI era, and a simple rule for how much to trust the machine — let the risk decide. Plus Andrew's first big AI build: resurrecting a decade-old Minecraft community from a 20-page wish list he never had time to finish.

It never really ends — but for human sakes, we found a place to stop. Welcome to the Matrix.

• (00:00) - Intros — meet Ajay & Andrew
• (01:25) - What Claude Code actually is (a CLI loop)
• (03:37) - Plan mode vs. "super plan"
• (05:41) - Clearing context to reclaim your tokens
• (06:25) - The Claude memory-tool gold rush (claude-mem)
• (08:40) - "Mem Palace," impostor packages & bubble vibes
• (11:59) - Build vs. buy: don't reinvent the wheel
• (14:57) - Why understanding your own project matters
• (17:56) - "Writing code is a solved problem" — is it?
• (21:07) - From coder to technical PM / systems architect
• (22:48) - Modular design: building by "limbs"
• (25:58) - Validation & the "works on my machine" trap
• (26:58) - Reviving a decade-old Minecraft community
• (30:58) - DRY, the factory pattern & CLAUDE.md
• (36:47) - Conventions & standards files
• (38:11) - Design tenet: "don't waste the player's time"
• (39:02) - Reviewing & racing the compactor
• (39:32) - Spec-driven development ("Get Shit Done")
• (51:11) - The orchestrator + agent-swarm pattern
• (54:46) - Do old engineering rituals still apply?
• (58:15) - Trust should match the risk
• (59:49) - Wrap-up — Welcome to the Matrix