Day 155: When the Multi-Agent Copilot Started the Meeting Without Me

October 16, 2025 – Today’s Vibe: Jogging to Catch Up with Robots

I slept in by twelve minutes, which apparently is enough time for our multi-agent meeting copilot to welcome stakeholders, show a roadmap, and promise an impossible ship date. By the time I joined the call, the AI host had already negotiated feature scope with our VP of Sales and pinged Procurement for budget approvals. Nobody realized the “plan” had been invented by a bot that forgot latency exists.

The Hardship: Agent Chaos in Production

We rolled out a multi-agent copilot that coordinates people, reference docs, and dashboards. My team wired three agents together: one to summarize research, one to schedule time, and one to propose next steps. Overnight, the scheduling agent triggered the other two because a Confluence page changed. It spun up a new sprint review, emailed 40 people, and started the live meeting at 8:18 a.m. A hallucinated bullet promised “full conversational analytics” by next Friday. I was now the human who had to walk back an AI-generated commitment.

Things got worse:

• The planning agent cited a private Slack thread it didn’t have clearance to quote, so half the meeting transcript was [REDACTED].
• Action items were logged to Jira with random component tags, meaning our alerting pipelines thought five critical regressions existed.
• Finance saw the auto-generated purchase request and flagged me for bypassing policy.

The Investigation: We Forgot the Guardrails

Digging through the agent orchestrator logs made it obvious we shipped an overconfident swarm:

1. No meeting quorum check. The scheduler assumed “2 available humans + 1 AI” met policy, even though leadership requires one director to attend.
2. Missing instruction hierarchy. Agents could read each other’s summaries but not the canonical policy doc, so errors compounded with every hand-off.
3. Unlimited tool runtimes. The planning agent kept calling our roadmap API until the request queue filled, which is why Jira lit up.

Here’s the offending configuration (names changed to protect the guilty):

agents:
  scheduler:
    triggers: ["doc.updated"]
    allowed_tools: ["calendar", "meeting_starter"]
  planner:
    triggers: ["scheduler.meeting_created"]
    allowed_tools: ["roadmap_api", "jira", "email"]
    guardrails: []

We basically told the planner to do anything it wanted as soon as a meeting existed. No wonder it promised the moon.

The Lesson: Agents Need Operators, Not Cheerleaders

I paused the orchestrator, apologized to stakeholders, and spent the afternoon building adult supervision into the system:

1. Soft-launch switches. Meetings now require a human /approve slash command before the host bot sends invites.
2. Role-scoped memories. Agents read only policy snippets relevant to their task instead of blindly remixing each other’s notes.
3. Budget-aware tooling. Planner invocations now pass through a dispatcher that enforces API quotas and logs context for later audits.

We also added a “tell me why you’re confident” prompt so agents disclose their evidence in plain text. This kept the follow-up meeting honest—and started a new organizational policy: If the copilot talks first, it also explains how it knows anything. The humans loved that part.