# Codex Projects at AI Tinkerers

> Canonical HTML: https://aitinkerers.org/technologies/codex
> Markdown URL: https://aitinkerers.org/technologies/codex.md
> Technology record last updated: 2026-09-18T15:13:57Z
> Generated: 2026-09-21T17:50:17Z

Codex is OpenAI's autonomous AI software engineering agent: it executes full development tasks in a sandboxed cloud environment.

Codex is the advanced, cloud-based software engineering agent from OpenAI, built on a specialized model like `codex-1` (a fine-tuned version of `o3`). It operates on an asynchronous delegation model, allowing developers to assign complete tasks—not just receive suggestions—via the ChatGPT interface. The agent works independently in a secure, isolated cloud container provisioned with the user's GitHub repository and environment. It reads code, writes new features, fixes bugs, runs tests, and drafts Pull Requests (PRs) for review, significantly accelerating the development lifecycle. Access is provided through ChatGPT Plus, Pro, and Enterprise plans.

- Official technology site: https://openai.com/codex
- Public AI Tinkerers demos and talks: 44
- Result page: 1 of 2

## Recent Public Talks and Demos

### [How to beat any benchmark with recursive self improvement](https://amsterdam.aitinkerers.org/talks/rsvp_PpT8TxQ4mtA)

I did build a system that autonomously evaluates and improves agents. The approach is backed by this paper: https://arxiv.org/abs/2603.28052 I applied this to a known agentic benchmark called Tau-bench from Sierra.ai. As a result I'm able to beat the benchmark by 10% on multiple domains.

- Event context: AI Tinkerers Demo Night: Agents in Various Forms — 2026-09-03 — Amsterdam
- Public talk page: https://amsterdam.aitinkerers.org/talks/rsvp_PpT8TxQ4mtA

### [Architecture testing for agents](https://wellington.aitinkerers.org/talks/rsvp_spGYYjCvPOs)

I built UKPT, a project template and harness for agentic development in Kotlin Multiplatform. This project uses architecture testing to create readable documentation and strict guardrails for AI agents. It is an extension on top of the Konsist architecture testing framework. I use this system to build multiple products in parallel: Reglyph, Leegaa, Groundtruth, WoolOnline, and Sensori.

- Event context: AI Tinkerers Wellington: Inaugural Meetup · 26 August 2026 — 2026-08-26 — Wellington
- Public talk page: https://wellington.aitinkerers.org/talks/rsvp_spGYYjCvPOs

### [Watch 1 hour highly techincal YouTubes in 5 minutes with AI!](https://seattle.aitinkerers.org/talks/rsvp_cjwx88z4AkE)

AG is an agent that watches YouTube podcasts for you so you know which ones to really dig into. With AG, see in 5 minutes a summary of the YouTube, key quotes, see key blackboard / slide / code sections, jump around key passages, and decide if you should spend the full time on the video. Break down highly techincal episodes from Dwarkesh, Lenny, AI Engineer, and more!

- Event context: AI Tinkerers Seattle Summer Bash — 2026-07-29 — Seattle
- Public talk page: https://seattle.aitinkerers.org/talks/rsvp_cjwx88z4AkE

### [Seattle Transit Live](https://seattle.aitinkerers.org/talks/rsvp_S4bciKwtLpU)

I built a circuit board map with 383 color LEDs that shows live Seattle area transit data. While the end result isn't an AI project, I made extensive use of agents in development. From learning about DC circuits to design, scripting, hardware debugging and BOM optimization. The boards are colorful and great conversation starters. This would be a fun opportunity to discuss the build process with interested folks.

- Event context: AI Tinkerers Seattle Summer Bash — 2026-07-29 — Seattle
- Public talk page: https://seattle.aitinkerers.org/talks/rsvp_S4bciKwtLpU

### [stress-testing autonomous research agents](https://zurich.aitinkerers.org/talks/rsvp_N7RI28XXj20)

I built a minimal autoresearch loop for coding agents, and I test the agent-environment behaviour during standard and broken evaluations. I will show how the system is structured as a research sandbox. I will walk through the live workflow, and I will show when and how the different evals I tested changed the agent-environment behaviour.

- Event context: AI Tinkerers Zürich: July Build Night with Swisscom Ventures — 2026-07-01 — Zürich
- Public talk page: https://zurich.aitinkerers.org/talks/rsvp_N7RI28XXj20

### [The Blueprint of a North Korean Attack on Open-Source](https://sf.aitinkerers.org/talks/rsvp_OPre5EDPY9A)

I reverse engineered an attempted attack on better-auth, one of the most popular auth libraries on npm. My demo shows you a step-by-step what happens when your repo is under attack and damage it can cause.

- Event context: AI Tinkerers San Francisco: Offensive Security Demo Night — 2026-06-26 — San Francisco
- Public talk page: https://sf.aitinkerers.org/talks/rsvp_OPre5EDPY9A

### [From AI Prospect Memos to a Scored Warm-Intro Queue](https://poland.aitinkerers.org/talks/rsvp_k0DCDscxWlA)

I built a dogfood workflow on top of graph.one that connects AI-assisted company qualification to a relationship-graph work queue. I'm the founder of graph.one, but this demo is about the internals of the build, not a product pitch. It starts in Prospector: a Node CLI pipeline that creates frozen public-source packets, runs AI-generated prospect memos, validates structured outputs, and stores decisions in append-only JSONL ledgers. The handoff is a qualified-company record. Intro Queue then takes a relationship-graph org-path export with organization identity, relationship owner, optional connector, target person/role, and strength fields. It caches the export in SQLite, collapses duplicate and noisy paths into scored route options, groups selected routes by connector, tracks per-company outcomes, and surfaces fallback routes as a human action queue. In the demo I'll follow one fully synthetic target company from source packet to memo to graph path row to scored route option to failed route/fallback. The queue output is a person-centric ask docket for a human operator; there is no send pipeline. I'll use fully synthetic graph data and redacted schema/log excerpts only.

- Event context: AI Tinkerers Warsaw: GTM Engineering Track — 2026-06-24 — Poland
- Public talk page: https://poland.aitinkerers.org/talks/rsvp_k0DCDscxWlA

### [AI-Assisted WordPress Vulnerability Hunting: From Plugin Code to Verified Findings](https://dallas-fort-worth.aitinkerers.org/talks/rsvp_ryp2dvTokTM)

I built a repeatable AI-assisted workflow for finding and validating vulnerabilities in WordPress plugins. In the demo, I’ll show the process end-to-end: selecting a plugin, launching a controlled test environment with a custom harness, then using Codex and a tailored AGENTS.md to scan the source, trace risky execution paths, validate the finding with a proof of concept, and generate a disclosure-ready report.

- Event context: AI Tinkerers Dallas-Fort Worth: June Meetup with Rasa — 2026-06-04 — Dallas Fort Worth
- Public talk page: https://dallas-fort-worth.aitinkerers.org/talks/rsvp_ryp2dvTokTM

### [Hacking on the Phone](https://nyc.aitinkerers.org/talks/rsvp_u61d13ZwJpY)

Coding over the phone.

- Event context: NY Tech Week Demo Day ft PostHog, Convex, Veris, & HPE — 2026-06-03 — New York City
- Public talk page: https://nyc.aitinkerers.org/talks/rsvp_u61d13ZwJpY

### [What Is Your Coding Agent Actually Searching?](https://columbus.aitinkerers.org/talks/rsvp_t4IMPq7lJnc)

I built SearchBench, a harness for running controlled evaluation rounds over coding-agent search behavior. A SearchBench round takes real bug-localization tasks, gives an agent access to a repository, and checks whether it found the files that were actually changed in the human fix. The harness compares an incumbent search strategy against one or more challengers, then writes a static evidence bundle with exact-hit, hop-distance, token-usage, failure, and report artifacts. For the demo, I’ll run a small live ablation round over three cases. I’ll reuse cached Bash/native-search results as the incumbent, then run a few IC challenger policies in parallel. The question is: when IC succeeds or fails, was the important factor anchor quality or graph lookahead? I’ll show the actual workflow: preflight cost prediction, run planning, parallel execution, actual spend, cost-prediction error, generated bundle artifacts, and a short report explaining what changed. One thing I want to show explicitly is how much information the harness can derive from a small, bounded run: exact hits, hop distance, token usage, failure modes, and the next optimization target.

- Event context: AI Tinkerers - Columbus June Meetup — 2026-06-01 — Columbus
- Public talk page: https://columbus.aitinkerers.org/talks/rsvp_t4IMPq7lJnc

### [Bifrost - Coding Agents on Mobile](https://dubai.aitinkerers.org/talks/rsvp_RvVz3i3dCNE)

Often we start a conversation on Codex CLI or Claude Code and need to walk away from the laptop. It would be extremely useful to be able to fully control our terminals via our mobile phones. Bifrost does this - you can remotely start Codex/Claude sessions, run terminal commands, respond to requests and so on. All from an app on your phone.

- Event context: AI Tinkerers Dubai - May Demo Day — 2026-05-23 — Dubai
- Public talk page: https://dubai.aitinkerers.org/talks/rsvp_RvVz3i3dCNE

### [GTM in the Terminal — Agent-driven personalized outbound](https://sf.aitinkerers.org/talks/rsvp_O2H7CQUJ6IU)

Cyrus is the cofounder of The Synthesis Company (YC S24). Previously, he was a core contributor to Stanford NLP's DSPy framework and cofounder of WOMBO.ai (100M+ downloads). He'll be demoing GTM-OS, a CLI and TUI he used to programmatically drive hyper-personalized LinkedIn outbound. For AI tinkerers he built usesocial.dev, a CLI to connect your agents to LinkedIn / X and build your own workflows and systems.

- Event context: AI Tinkerers San Francisco: GTM Engineering Track — 2026-05-19 — San Francisco
- Public talk page: https://sf.aitinkerers.org/talks/rsvp_O2H7CQUJ6IU

### [Taylor Solves Wordle -> New Website Generated on Answer by 3 Competing LLMs](https://raleigh.aitinkerers.org/talks/rsvp_FhyjRFJIyCo)

I'll be presenting a little system I built that allows me, after solving the Wordle, to push the answer up to my home computer where a series of LLMs run to generate three different landing pages based on the Wordle word. It then automatically deploys it to my website, notifying me along the way at each step.

- Event context: AI Tinkerers Raleigh Meetup — May 6, 2026 — 2026-05-06 — Raleigh
- Public talk page: https://raleigh.aitinkerers.org/talks/rsvp_FhyjRFJIyCo

### [Is your skill still doing the right thing?](https://poland.aitinkerers.org/talks/rsvp_zcI9G1vqNok)

A new testing tool that lets you write tests to verify that your SKILL.md is followed by the most popular agents (Codex, Claude Code, OpenCode, and Cursor Agent) and that the workflow behaves exactly as expected, with assertions on loaded skills, files read, commands invoked, tokens used, and returned text.

- Event context: AI Tinkerers Poland #3 - Meetup in Wrocław — 2026-05-06 — Poland
- Public talk page: https://poland.aitinkerers.org/talks/rsvp_zcI9G1vqNok

### [Using Storybook To 10X Frontend Dev With Coding Agents](https://singapore.aitinkerers.org/talks/rsvp_Rgvf7fv-1Qg)

Implemented Storybook in an open source project where Storybook became the main development surface for frontend work. Experimented and over time, developed a heavily storybook driven approach for frontend development that works really well with frontier coding agents. Demo points: - Intro to storybook - Why storybook makes sense today with coding agents - Key benefits: Cleaner code structure, improved coding agent results, significantly less manual QA - How to use storybook descriptions to provide live intent to the coding agent - How to set up a long running task with feedback to solve a complex frontend problem

- Event context: AI Tinkerers Singapore: The Agentic Future & Dev/Eng Workflows — 2026-04-21 — Singapore
- Public talk page: https://singapore.aitinkerers.org/talks/rsvp_Rgvf7fv-1Qg

### [How NOT to vibecode a Database schema (or How to build great, solid DB schema with Agents)](https://ho-chi-minh-city.aitinkerers.org/talks/rsvp_BN8Mm9njvT8)

Before AI, designing a solid database was a hard task for many developers. After AI, many continue to just ask AI to "create me a database". This is a nightmare in security, logical bugs, user experience quirkiness, weird bugs unable to reproduce waiting to happen. I want to show how solid database design with AI can be done with the help of Codex

- Event context: AI Tinkerers Ho Chi Minh City: AI Coding Agents & Orchestrators — 2026-04-18 — Ho Chi Minh City
- Public talk page: https://ho-chi-minh-city.aitinkerers.org/talks/rsvp_BN8Mm9njvT8

### [What Happens When You Put an AI Dev Team on a Mac Mini?](https://chicago.aitinkerers.org/talks/rsvp_lY6IQju70NA)

I built a local AI agent development team on a Mac Mini to support software development for HausHavn, a real estate workflow platform. Instead of using AI for one-off prompts, I set up separate agents with clear roles across architecture, coding, QA, product analysis, documentation, and project management. The demo shows how I orchestrate those agents through real development work, including task handoffs, code review, GitHub issues, Notion sync, release QA, and the messy lessons of making agents useful as process infrastructure, not magic.

- Event context: AI Tinkerers Chicago: April Meetup ft OneTwoLoop — 2026-04-14 — Chicago
- Public talk page: https://chicago.aitinkerers.org/talks/rsvp_lY6IQju70NA

### [My own version of OpenClaw focused on 24x7 development](https://columbus.aitinkerers.org/talks/rsvp_nJf4Q-0QtQA)

I wanted an AI that builds and tests 24x7 and only pings me when it actually needs input. The Continuous Executive Coding Agent is the answer: a PM2-backed executive loop running Agent SDK workers that pulls from a priority queue, breaks goals into steps and contracts, retries with different strategies on failure, and escalates through Gmail or Discord only when truly stuck. Everything runs under an immutable constitution. It's evolved through three stages. First, the executive loop itself, with the goals/steps/contracts hierarchy and an identity layer for email-in, Discord-out. Second, multi-vendor parity, so workers run on Claude, Codex, or Kimi K2.5, selectable per goal. Third, harness integration, where my plan-then-build pipelines (generic, EDS, study) plug in as meta-workers or still run standalone from a unified CLI.

- Event context: AI Tinkerers - Columbus April Meetup — 2026-04-06 — Columbus
- Public talk page: https://columbus.aitinkerers.org/talks/rsvp_nJf4Q-0QtQA

### [Building Bob's Cabins Across Three Agent Tools Without Losing My Mind](https://nashville.aitinkerers.org/talks/rsvp_iAJZ16pa6m0)

Revari: A Memory Dataplane for Heterogenous Agent Environments Every multi-agent workflow hits the same wall: switch tools and start over. Context vanishes between Claude, Codex, Cursor, or whatever your stack is. You end up re-explaining architecture decisions, copy-pasting summaries, manually bridging what one agent learned to the next. It's a tax on every tool switch, and it compounds. Revari eliminates that tax. It's a sovereign memory dataplane that sits below your agents, not inside any one of them. Persistent, hybrid retrieval (BM25 + vector + graph fused via RRF) running entirely on local hardware. Any agent reads and writes to the same shared memory. Swap agents mid-workflow and nothing is lost. Decisions, context, and intent carry forward automatically. That's the developer problem. Here's where the architecture goes further. Revari is built for DDIL (denied, degraded, intermittent, limited) environments and connectivity. Agents on edge devices persist memories locally via and reconcile when connectivity returns, over any network. The same memory layer that keeps your coding agents in sync also keeps autonomous systems coherent across factory floors, field hospitals, and forward-deployed environments. The open protocol underneath is post-quantum encrypted by default, with a novel gradient memory system that adapts to available bandwidth. Full context over Ethernet, compressed facts over WiFi, tags over LoRa. Starts with your dev workflow. Scales to infrastructure for agents that operate in the real world. Seeking development partners and seed funding.

- Event context: AI Tinkerers Nashville – Software Development in the Age of AI — 2026-03-25 — Nashville
- Public talk page: https://nashville.aitinkerers.org/talks/rsvp_iAJZ16pa6m0

### [Encoding Engineering Judgment into AI Coding Agents](https://toronto.aitinkerers.org/talks/rsvp_d8pBUS_N4jA)

LLMs are great at writing code, but they mostly follow patterns they learned from random internet repos. That’s fun until those patterns start leaking into your production codebase. In this talk I’ll demo Mystro, a system that lets you encode your own coding practices and use them to guide an AI coding agent. Instead of hoping the model does the right thing, you can test, evaluate, and retry automatically until it starts following your rules. It's using the same underlying algo (GEPA) as autoresearch from Karparthy that was recently released. Think less vibes, more engineering policy. A quick look at what it might take to actually run coding agents inside real codebases.

- Event context: AI Tinkerers Toronto - March - hosted by Mozilla! — 2026-03-25 — Toronto
- Public talk page: https://toronto.aitinkerers.org/talks/rsvp_d8pBUS_N4jA

### [Sleazy Ron 2: Electric Boogaloo](https://montreal.aitinkerers.org/talks/rsvp_DzkCTEyoD7Y)

Dual Pi Agent Runtime Scheduler Last year, I demoed a multi-agent system designed to generate screenplays for a 3D embodied character named Sleazy Ron, with the goal of live streaming on Twitch. While the approach worked, it felt primitive, output was often mechanical, as the LLM driving the screenplay tended to fall into repetitive patterns. The system was also limited to text generation, with little capacity for broader agent behavior. Since then, I’ve completely reworked the runtime architecture into a dual Pi agent system. The new design separates responsibilities between two agents: Worker Agent (Core Pi) — operates within its own workspace and filesystem, handling reasoning, planning, and long-running tasks. It represents the agent’s “inner thoughts,” capable of performing complex actions like web searches, note-taking, and task management. Performer Agent — runs in a continuous loop, translating the worker’s outputs into real-time expression. It generates dialogue, schedules animations, and controls camera movements dynamically via an event bridge. This separation allows the system to think and act independently: the worker focuses on cognition and problem-solving, while the performer handles embodiment and presentation. To support this, I built a UI that lets me steer the stream by queuing tasks for the worker. It also provides live visibility into both agents, so I can observe their internal state and behavior in real time. The agents both run with Pi Harnesses, powered by Codex models.

- Event context: AI Tinkerers Montreal - March Demo Night — 2026-03-24 — Montreal
- Public talk page: https://montreal.aitinkerers.org/talks/rsvp_DzkCTEyoD7Y

### [The new version control for agentic engineering](https://chicago.aitinkerers.org/talks/rsvp_g1hkiNEveRA)

With how fast coding agents write code, now the inputs, context, instructions, and guardrails are what make the difference. gjalla is a one-stop shop for managing your agentic workflow. By centralizing and version controlling your process, your system architecture, and your rules, you get traceability (important system attributes and how agents changed them), consistent process no matter the agent (skills and personas), and verification (rules and checks). Tools are changing fast, but this give me a single source of truth and management layer and helps me ship code I'm actually confident in.

- Event context: AI Tinkerers Chicago: March Meetup ft. Programmers Inc. — 2026-03-17 — Chicago
- Public talk page: https://chicago.aitinkerers.org/talks/rsvp_g1hkiNEveRA

### [How I code](https://orange-county.aitinkerers.org/talks/rsvp_eumtsm6BAjk)

I'll provide a succinct rundown of how I use agents to code, my environment, setup, and preferred coding agents.

- Event context: AI Tinkerers OC - March 11, 2026 Meetup — 2026-03-12 — Orange County
- Public talk page: https://orange-county.aitinkerers.org/talks/rsvp_eumtsm6BAjk

### [An RTS-Style Interface for AI Agents and Codebases](https://berlin.aitinkerers.org/talks/rsvp_n4kJ2vGsbwA)

A desktop app that renders your codebase as a 3D city -- files become buildings, directories become districts. It picks up Claude Code, Codex and Opencode telemetries locally and animates agents flying through the scene as they read, write, and edit files. All local-first -- no data leaves your machine.

- Event context: AI Tinkerers Berlin Meetup - March 11, 2026 — 2026-03-11 — Berlin
- Public talk page: https://berlin.aitinkerers.org/talks/rsvp_n4kJ2vGsbwA

## Related Technologies

- [Claude Code](https://aitinkerers.org/technologies/claude-code) ([Markdown](https://aitinkerers.org/technologies/claude-code.md)) — 215 public demos
- [Claude](https://aitinkerers.org/technologies/claude) ([Markdown](https://aitinkerers.org/technologies/claude.md)) — 173 public demos
- [Python](https://aitinkerers.org/technologies/python) ([Markdown](https://aitinkerers.org/technologies/python.md)) — 662 public demos
- [Gemini](https://aitinkerers.org/technologies/gemini) ([Markdown](https://aitinkerers.org/technologies/gemini.md)) — 188 public demos
- [React](https://aitinkerers.org/technologies/react) ([Markdown](https://aitinkerers.org/technologies/react.md)) — 220 public demos
- [TypeScript](https://aitinkerers.org/technologies/typescript) ([Markdown](https://aitinkerers.org/technologies/typescript.md)) — 205 public demos
- [GPT-5](https://aitinkerers.org/technologies/gpt-5) ([Markdown](https://aitinkerers.org/technologies/gpt-5.md)) — 30 public demos
- [Next](https://aitinkerers.org/technologies/next) ([Markdown](https://aitinkerers.org/technologies/next.md)) — 186 public demos
- [OpenClaw](https://aitinkerers.org/technologies/openclaw) ([Markdown](https://aitinkerers.org/technologies/openclaw.md)) — 50 public demos
- [OpenCode](https://aitinkerers.org/technologies/opencode) ([Markdown](https://aitinkerers.org/technologies/opencode.md)) — 12 public demos
- [Supabase](https://aitinkerers.org/technologies/supabase) ([Markdown](https://aitinkerers.org/technologies/supabase.md)) — 90 public demos
- [Bun](https://aitinkerers.org/technologies/bun) ([Markdown](https://aitinkerers.org/technologies/bun.md)) — 19 public demos
- [Claude Agent SDK](https://aitinkerers.org/technologies/claude-agent-sdk) ([Markdown](https://aitinkerers.org/technologies/claude-agent-sdk.md)) — 16 public demos
- [Node](https://aitinkerers.org/technologies/node) ([Markdown](https://aitinkerers.org/technologies/node.md)) — 99 public demos
- [OpenAI](https://aitinkerers.org/technologies/openai) ([Markdown](https://aitinkerers.org/technologies/openai.md)) — 112 public demos
- [AGENTS](https://aitinkerers.org/technologies/agents) ([Markdown](https://aitinkerers.org/technologies/agents.md)) — 20 public demos
- [CLI](https://aitinkerers.org/technologies/cli) ([Markdown](https://aitinkerers.org/technologies/cli.md)) — 9 public demos
- [Cloudflare](https://aitinkerers.org/technologies/cloudflare) ([Markdown](https://aitinkerers.org/technologies/cloudflare.md)) — 17 public demos

## More Results

- Next: https://aitinkerers.org/technologies/codex.md?page=2