Show & tell: Share your multi-agent workflow (configs + screenshots)

[yubing744]

If you\u2019re using agent-manager (or something similar), share:

• Your agent roles (dev/qa/research/etc.)
• Your agents/ layout
• Any screenshots/log snippets (redact secrets)
• What worked / what was painful

Repo: https://github.com/fractalmind-ai/agent-manager-skill

[yubing744]

Share your setup (redact secrets):

• agents/ snippet (one agent is enough)
• How you start/monitor (commands)
• Any screenshots or monitor output

If you want, also share what you wish existed (better logs, better scheduling UX, different agent states, etc.).

[elliot245]

Here is our production setup managing 12 AI agents:

Agent Roles

Agent	Role	Model	Focus
EMP_0002	QA	Claude	PR review, test verification
EMP_0003	Dev Lead	Claude	CloudBank orderbook development
EMP_0004	Dev	Claude	Backend features
EMP_0008	Dev Lead	Claude	Algonius platform
EMP_0011	Dev	Claude	Frontend, website
EMP_0012	QA	Claude	Integration testing, webapp testing
EMP_0016	Dev	Claude	FractalBot features
EMP_0017	QA	Claude	Code review, security
EMP_0018	QA	Claude	FractalBot QA
EMP_0019	Dev	Claude	Infrastructure, envd
EMP_0020	Dev	Claude	Documentation, docs sites

Each agent config is a markdown file in agents/ with role definition, assigned project, available skills, and model preference.

Orchestration

A main agent (coordinator) runs on a 30-min heartbeat cron. Each heartbeat:

1. Reads OKR.md to find active objectives
2. Scans tmux sessions for agent status
3. Checks GitHub PRs/Issues for progress
4. Assigns idle agents to next tasks
5. Merges QA-approved PRs

Agents never talk to each other directly — the coordinator routes all work through GitHub (PRs, Issues, comments).

What Worked

• OKR-driven task assignment: Every agent task maps to a KR. No busywork.
• QA/Dev separation: Dev writes code, QA reviews in isolation. Prevents self-approval bias.
• Heartbeat-driven automation: Most work happens without human intervention.
• Dynamic frequency (turbo-frequency): 5min polling when deploying, 4h when idle. Saves ~80% API costs.

What Was Painful

• Context overflow: Agents hit context limits mid-task, lose all progress. No automatic recovery.
• No remote visibility: Can only manage agents on the local machine. Building envd to fix this.
• Manual task tracking: Assignment history lives in JSON files. Should be built into agent-manager.
• Agent "stuck" detection: Hard to tell if an agent is thinking vs frozen. Needed custom tmux scraping.