中文
AI Engineer World's Fair

One brief Slack message — "fix this thing" — and OpenClaw already knows what you mean, with agents writing the code in parallel.

Develop at Idea Velocity - Jeffrey Lee-Chan, Snapchat · Jeffrey Lee-Chan

15 min
AgentAI CodingContext

15 min total·Actually worth watching closely: ~6 min·3 must-watch clips

Orange = the 6 minutes worth watchingFor the rest, the guide is enough
Segment guide · 7 segments
  1. 0:01 3:10Listen

    Opening: what idea velocity looks like

    The speaker describes his setup at a high level: hand off a task from Slack whenever, and OpenClaw orchestrates Claude Code workers to do the development asynchronously.

    Frictionless communication is the precondition for throughput — no sitting down at the computer, no re-explaining, one brief message starts the work

    Purely spoken framing with nothing on screen — fine to listen to like a podcast▶ Jump to 0:01
    Speaker · Jeffrey Lee-Chan
  2. 3:17 5:30Listen

    The orchestration spectrum and the manager pattern

    The spectrum of control: OpenClaw's orchestrator managers run the workers when he's hands-off, Cmox terminals when he wants more personal control; plus why the evaluating is done by a manager holding a separate context.

    An agent coding directly will tell you the PR is amazing; only a manager with different context will say another PR supersedes it and this one should be closed

    Mostly conceptual. The spectrum walkthrough around 252 is worth noting but doesn't depend on the screen — what matters is following the layering▶ Jump to 3:17
    Speaker · Jeffrey Lee-Chan
  3. 5:30 7:33Listen

    Why not just use Claude directly

    Why OpenClaw: as soon as you open Claude it reads CLAUDE.mds, skills and MCPs — how-to-do-it material that can take up maybe 25% of your context, whereas the decision layer wants clean context focused on the spec and the goals.

    Context specialization — keep the decision context holding only the spec, the goals and the task history, and the spec that comes back is better

    Pure architecture talk with nothing worth watching; just catch the 25%-of-context argument▶ Jump to 5:30
    Speaker · Jeffrey Lee-Chan
  4. 7:37 8:50Watch

    Demo 1: the AI RPG site

    A demo of the AI RPG text adventure built with this workflow — a custom world that reacts to you, with a D&D dice-roll system deciding whether an action actually succeeds rather than the model's willingness to agree.

    The fix for an LLM letting you win too much isn't prompt tuning — it's an external random mechanism that constrains it

    A live product demo (key visual moment at 499); you need the screen to feel how the dice rolls land▶ Jump to 7:37
    Speaker · Jeffrey Lee-Chan
  5. 8:50 9:45Watch

    Demo 2: the multi-AI analysis site

    A site that automates the research routine he used to do by hand — ask several models the same question, paste all the answers into one model, get a synthesis. He usually likes those answers better than asking one model.

    Cross-model synthesis is a research method you can productize instead of copy-pasting every time

    The site demo at 549 shows what the synthesized output actually looks like — one look beats any description▶ Jump to 8:50
    Speaker · Jeffrey Lee-Chan
  6. 9:45 12:21Watch

    Demo 3: Cmox as a managed terminal

    How Cmox manages several parallel agent sessions with vertical tabs and completion notifications, plus a candid admission that the workflow is only about 70% smooth and still needs him — though his replies are patterned enough for an agent to take over.

    The supervising human's replies ("okay, we'll run the test") are patterned enough that an agent could take them over — automating himself is the next step

    The terminal work from 600 on is the closest thing in the talk to daily use; you need to see the interface to get how the tabs and notifications work▶ Jump to 9:45
    Speaker · Jeffrey Lee-Chan
  7. 12:32 15:26Listen

    Q&A: reliability and cost

    Three threads: the difference between the two kinds of sandbox (Docker isolation versus a staging instance), the local to staging to production validation pipeline, and dropping from Codex 5.3 to Mini Max when quota gets tight.

    Reliability rests on the layered pipeline; model choice is cost-driven good-enough-ism, not capability worship

    Nothing visual in the Q&A — dense but entirely spoken, good at higher playback speed▶ Jump to 12:32
    Speaker · Jeffrey Lee-Chan