Skip to content

LIVE-RUN.md

Latest commit

 

History

History
91 lines (68 loc) · 4.73 KB

LIVE-RUN.md

File metadata and controls

91 lines (68 loc) · 4.73 KB
Raw

The live run — rules and runbook

One of the three pitch slots can be a REAL agent working live, if and only if it earns the slot in rehearsal. This document is the complete set of rules for making that safe.

The one rule that decides everything

Rehearse the live run twice, complete, before the pitch. Two clean passes → it takes slot 3 (replacing the recorded land-pickup). One failure → it stays recorded and the live run becomes the Q&A encore. Decide from evidence, not optimism, and decide BEFORE walking on stage.

What "live" means here

operator sentence → agent (Claude / Codex / ChatGPT with the skills) → MCP → usedonald.com/runs/<key>

The agent is any MCP-capable runner. Donald never runs agents; it watches them. That is the thesis — "the interface builds itself from natural language" — and it is also why this demo is honest: the judge watches the same pipeline any customer would use.

Pre-flight (do all of it the same morning)

  1. Agent session open BEFORE going on stage, with BOTH skills loaded: skill/donald-flow/SKILL.md and skill/nauta-operations/SKILL.md, and the MCP connected: { "mcpServers": { "donald": { "type": "http", "url": "https://mcp.usedonald.com/v1/mcp" } } }
  2. Smoke run that morning: one tiny run (start_run → one action → finish_run). Confirm it appears at usedonald.com/runs/<its-key>. If the smoke fails, the live slot is OFF — no debate.
  3. Fourth browser tab already open on the agreed run key (see below). The tab shows CONNECTING until the run starts — that is normal and even good theater.
  4. wait tool present in the agent's tool list (pacing on). If it is missing, the run finishes in seconds and the demo is dead — check it in the smoke run.
  5. Laptop on power, sleep disabled, next start + tunnel + both containers up (they auto-restart, but verify: docker ps shows donald-api and donald-mysql).

The run key convention

live-<event>-<n> — e.g. live-pitch-001, live-judge-001. Agreed before the pitch, typed into the fourth tab beforehand. Never improvised on stage. Keys are per-client and case-sensitive.

The super prompt (paste as-is, fill the blanks)

You are Nina, Nauta's operations agent, working for Mueblerías Berríos.
Load and follow the donald-flow skill exactly: report every step to Donald AS IT HAPPENS,
never batch at the end. Use run_key "live-pitch-001". Show the watch_url first.

Use the wait tool to give every step its declared duration - a step that says 20s takes 20s,
with report_progress between waits. The duration is the intervention window.

Declare a plan of 5-7 steps maximum. Narrate findings with real identifiers from the
nauta-operations world. Before any irreversible or costly action, block on a decision with
2-3 priced options. Poll check_instructions every ~10s while blocked; honour whatever the
operator picks, then resolve_instruction and continue. Finish with finish_run and a money
figure in the summary.

The request: <THE JUDGE'S OR OPERATOR'S SENTENCE GOES HERE>

Reserve request, if the judge does not offer one: "L'Oréal's invoice does not match our PO. Find out what happened and draft the right email."

Stage choreography (the 90-second slot)

t Screen Say
0:00 Agent window: paste the sentence "This is the whole integration: one sentence."
0:10 Fourth tab: the plan lands in grey "Nobody wrote this screen. The agent proposed it."
0:30 Steps light and narrate Read one finding aloud, with its identifier.
0:55 The gate opens with priced options Click one ON the projector — the judge sees the write path.
1:20 The agent resumes, finishes "That instruction went back over the same protocol."

If the agent is slower than this, talk over it — the amber wait IS the product. If it stalls past ~30s with nothing moving, say "and while it works, let me show you the same flow recorded" and switch to the land-pickup tab. No apology, no fiddling.

Failure playbook

Symptom Do
Run never appears in the tab Switch to recorded land-pickup, keep talking. Encore dead.
Agent skips the gate Let it finish; the run is still a live graph — narrate that.
Agent finishes too fast (no wait tool) Same switch. Verify pacing in the NEXT smoke run.
Tunnel dies mid-run localhost:3000 mirrors everything — swap the tab, carry on.

What we do NOT do

  • No agent-runner inside Donald, no textbox that spawns agents. The agent is external on purpose.
  • No improvised run keys, no untested judges' hardware, no live run without the morning smoke.
  • No specialized task agents (email readers, PDF parsers) before the pitch — that is the roadmap slide, not today's build. The skill already supports them without touching Donald.