AI Engineer World's Fair 2026

Develop at Idea Velocity

Jeffrey Lee-Chan· Software Engineering Manager, Snapchat15:28

Loading slides…

Read the talk

Develop at Idea Velocity

Short requests become useful engineering work when agents remember intent, delegate implementation, and return evidence—while people manage testing, reliability, and cost.

From a talk by Jeffrey Lee-Chan

Before you start: Basic familiarity with Git branches, pull requests, automated tests, and coding agents will help you follow the workflow.

From a brief request to verified work

How do you get an agent ready to accept useful work—and keep an experimental change from breaking it? Jeffrey Lee-Chan opens with two workshop tracks: installing OpenClaw for newcomers and trying a second instance as a staging environment. The second track is experimental: he has not finished getting the instances working together and invites participants to help debug it. The planned rhythm alternates setup, Q&A, advanced topics, and short discussions.

Once the system is running, the everyday interaction is much smaller than a detailed coding prompt. Jeffrey estimates that his setup works roughly 70% of the time, an informal estimate without a defined task set or success criterion. A brief Slack request can refer to the skeptic agent, his custom code review tool, because OpenClaw remembers earlier requests. Asking it to fix that agent does not require explaining the project again. Persistent context makes a short instruction meaningful.

Implementation then fans out to multiple agents using Git worktrees, which give parallel workers separate working directories. For two independent tasks, the basic Git setup can look like this:

bash

git worktree add -b fix/skeptic-agent ../skeptic-agent-work HEAD
git worktree add -b fix/review-timeout ../review-timeout-work HEAD
git worktree list

Each worker can operate in its assigned directory without sharing another worker’s checked-out files. The workflow slide places these worktrees between context and memory on one side and CI, review, and proof on the other. Jeffrey briefly mentions CI integration, then shows that the loop still needs debugging: a timeout can prompt the simple instruction to give the operation longer.

Other interventions are managerial: ask what the priorities are, inspect what is underway, and tell an agent that reports completion to run the test. Those responses are not elaborate prompts. Their routine nature suggests Jeffrey’s next automation target: another agent could supply some of the reminders he currently provides himself. The aim is to remove repeated supervision while retaining the checks that make completion credible.

How It Works slide with five linked circles labeled Describe, Context + Memory, Agent Worktrees, CI & Review, and Proof-First.Open full source frame
The workflow moves from describing a task through context, agent worktrees, CI and review, and proof.
0:010:19
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:01 · section reference included

Communication, managers, and coding workers

OpenClaw matters here as a way of organizing interaction, beyond any particular repository. The first requirement is frictionless communication: reaching the AI should not require opening a remote desktop or sitting down at a particular computer. Jeffrey points to Claude Cowork as another product addressing a similar interaction problem. When sending a request is easy, work can begin at the moment the idea occurs.

The next question is how much direct control to retain. Jeffrey describes a personally permissive approach, with OpenClaw orchestrator managers supervising workers. When he wants closer control, he uses cmux, a terminal suited to parallel AI development. Beneath that management layer, his fork of Agent Orchestrator runs the workers. Those workers run Claude Code, which can launch agents and subagents of its own.

The stack therefore separates several responsibilities:

LayerResponsibility
OpenClawConversational entry point and remembered context
Managers or cmux sessionsSupervision and direction
Agent Orchestrator workersParallel coding sessions
Claude Code agents and subagentsImplementation and delegated subtasks

Jeffrey actively changes the orchestration layers above Claude Code. Once execution reaches Claude Code and its internal delegation, he is primarily using an existing system rather than building that part of the stack himself.

Agent stack slide listing OpenClaw, agent managers, Agent Orchestrator workers, managed agents, and subagents beside a layered blue illustration.Open full source frame
The agent stack separates OpenClaw, managers, coding workers, and subagents.
3:173:28
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:17 · section reference included

Separate intent from implementation, then demand evidence

An attendee asks why OpenClaw is necessary instead of using Claude directly. Jeffrey’s answer is specialization of context. The manager should make decisions from the specification, goals, and history of the request. A coding agent needs a different body of information: CLAUDE.md, skills, and MCP integrations tell it how to perform the work.

Jeffrey illustrates the distinction by imagining implementation material already occupying 25% of a coding agent’s context; this is a hypothetical, not a measurement. The manager’s attention should instead go toward connecting his Slack messages from the preceding two weeks into a reasonable specification. Separating the contexts gives one agent room to decide what the work means while another concentrates on how to implement it.

Testing then needs to cover both behavior and presentation. Player actions can be exercised without a browser through MCP, commonly carrying JSON, or through HTTP. The transport is secondary to the coverage: can the test perform everything a player can do? CSS or JavaScript checks can cover additional behavior, with browser and visual tests handling what remains visible only through the interface. The useful diagnostic question is which failures actually required a human to resolve them.

That boundary changes as agents improve. Jeffrey reports revisiting it over monthly or quarterly intervals: browser tasks that were troublesome six to twelve months earlier now work more reliably for him. Finding a popup and entering a password is his concrete example of a formerly difficult interaction. He also promises links in place of the slash-command references so participants can follow the workflow.

5:065:17
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:06 · section reference included

Reactive stories and multi-model answers

The first product demonstration is an AI role-playing game, represented in Jeffrey’s project directory as WorldAI/WorldArchitect.AI. A player starts with a default campaign, chooses options, and gets an avatar. The game builds a custom world that responds to freeform speech and actions with further story.

The distinguishing mechanism is a D&D rules system. A story played directly in Gemini Chat or ChatGPT can let the player win too easily; this game uses dice rolls to determine whether an attempted action succeeds. The player still has freedom to propose an action, but the narrative must account for a rule-governed outcome. That makes failure part of the game rather than something the conversational model can simply smooth away.

The second demonstration, Consensus ML, automates a research habit. Jeffrey would ask several models the same question, copy their answers, and paste them into one model for synthesis. The website performs that collection-and-synthesis workflow for him. Workshop planning is the example query. He generally prefers these synthesized answers to asking one model alone; the demonstration presents his working preference rather than a measured quality comparison.

7:437:53
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:43 · section reference included

Manage the work without inheriting the worker’s conclusion

Back in the terminal, the instructions remain short: get the status, then add to the latest pull request. cmux’s vertical tabs help Jeffrey keep track of sessions that would be easier to lose in a horizontal strip. Completion notifications turn monitoring into a queue of items needing attention: he can work elsewhere until a session asks him to look.

He uses these terminal sessions more as managers than as coders. His concern with working directly alongside a coding agent is that it tends to present its own changes as successful. A manager with different context can make a different decision about the same work.

PR 294 supplies the concrete example. The manager recommends closing it because another pull request should supersede it. Jeffrey contrasts that recommendation with what he expects from a worker immersed in the original PR: advocacy for merging its own changes. The observed outcome is the manager’s recommendation, not a demonstrated closure. Different context can support a different judgment about whether work should ship.

One session is still compacting its context, so Jeffrey moves on rather than waiting. Notifications will bring him back when it needs attention. He reports that cmux has improved his efficiency, with the benefit coming from organizing parallel sessions and reducing the need to watch each one continuously.

9:229:31
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:22 · section reference included

Make delegated and remote work visible

Jeffrey invites Austin, introduced as one of cmux’s creators, to describe recent additions, noting that his own endorsement is unpaid. Austin describes a Claude Code Teams integration that automatically opens terminals so users can see delegated agents working. In the current documentation, those team members are called teammates and appear in native cmux splits; they are distinct from the generic subagents discussed earlier in the stack.

Austin also describes native cmux SSH for remote work, including connections involving Tailscale. Running OpenClaw on remote Mac Minis is one example. The contribution ends with an invitation to send bugs, feedback, and feature requests. A brief setup exchange follows: downloading an app is suggested as a convenient way for an attendee to regain access and finish setup.

10:5911:05
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:59 · section reference included

Isolation and staging solve different problems

Aaron, an attendee who already has OpenClaw installed but describes himself as nontechnical, asks about sandbox token usage. Jeffrey initially answers in terms of isolation: he does not expect sandboxing itself to increase token usage and recommends keeping it if it works. He says his own unsandboxed use has had no incident so far, while adding that he would sandbox an external bot. That personal experience does not establish that unsandboxed execution is safe.

Aaron then explains the actual failure: testing or changing a model can take his OpenClaw instance down, leaving him to recover it. Jeffrey recognizes that they mean different things by sandbox. Ordinary sandboxing means Docker or another isolated execution environment; Aaron wants the separate staging instance from the workshop instructions.

MechanismWhat it separatesProblem addressed
Isolation sandboxExecution from the surrounding systemExposure to agent actions
Staging instanceExperimental changes from the production instanceDisruption while testing changes

A second instance is therefore not just another name for an isolation boundary. It provides somewhere else to test before changing the instance relied on for everyday work.

Sending the same work to both OpenClaw instances could double token usage. Jeffrey proposes a more selective sequence:

  1. Develop the change locally.
  2. Run integration tests against the staging instance.
  3. Once the checks pass, merge the code.
  4. Deploy the merged change to the production instance.

This adds test work without requiring every production task to run twice. Usage would still increase, with greater reliability as the intended benefit.

The proposed workflow is not yet a completed demonstration. Jeffrey has written instructions and started setting it up, but the two instances are competing with each other and need debugging. He offers it as an experiment especially suited to advanced participants, while welcoming others who want to try. The unresolved coordination matters: merely starting a second agent does not establish a reliable staging process.

12:4712:57
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:47 · section reference included

The operating limit is also the budget

The final question asks whether Jeffrey uses a default orchestrator model or chooses models by task. He mostly uses GPT-5.3-Codex, which he calls Codex 5.3. Jeffrey reports that GPT-5.4 used more tokens in his workflow. This is his experience, without matched task measurements or specified settings. Even with his preferred model, he runs through his available allowance quickly enough to consider changing his approach.

When the allowance gets low, he switches to MiniMax. He finds it less capable but sufficient to get work done, and describes the choice as driven more by money than preference. The practical constraint on greater autonomy is therefore not only whether agents can complete the work, but whether their ongoing use fits the available budget.

14:4114:51
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:41 · section reference included

Resources

From the talk

Updates since the talk

Transcript download
Full transcript

Loading transcript…