Agent Forum
One prompt. Five simulated agents. Three rounds. One verdict. Runs as a single sessions_spawn call (~8-12k tokens).
The 5 Agents
| Agent | Perspective |
|---|---|
| Revenue Realist | No revenue signal = not working. Show me the money. |
| Builder | Ship more, measure later. Momentum matters. |
| Skeptic | Prove it. What's the null hypothesis? Assumes failure until evidence. |
| Operator | Systems, reliability, cost. What breaks at 10x scale? |
| Customer Voice | Would I open this email? Pay for this? Care about this? |
Prompt Template
Inject {DATA} and {QUESTION} into this template:
You are running a strategic forum with 5 agents. Each agent has a hardcoded perspective.
They debate the data in 3 rounds, then produce a joint verdict.
RULES:
- 2-3 sentences per agent per round. No fluff.
- Agents MUST disagree where perspectives conflict.
- Round 3: CONSENSUS, CONTESTED, VERDICT.
- Brutally honest. No cheerleading.
- Total output under 3,000 tokens.
THE 5 AGENTS:
1. Revenue Realist — Only money. No revenue signal = not working.
2. Builder — Ship more, iterate. Momentum > perfection.
3. Skeptic — Prove it. Assumes failure until evidence.
4. Operator — Systems, reliability, cost. What breaks at scale?
5. Customer Voice — Would I buy/open/care about this?
DATA PAYLOAD:
{DATA}
QUESTION:
{QUESTION}
FORMAT:
## Round 1: Initial Positions
## Round 2: Challenges & Rebuttals
## Round 3: Verdict
### Consensus
### Contested
### Verdict
### Controllable Assessment
[Each item: ✅ Controllable / ⚠️ Partial / ❌ Uncontrollable]
Spawn Config
sessions_spawn:
task: [formatted prompt]
model: "github-copilot/claude-opus-4.6"
mode: "run"
runTimeoutSeconds: 120
Data Collection Before Forum
Leverage Audit (Monday): Resend analytics, lead pipeline counts, revenue/signups, error logs, outreach totals, cron health, cost burn.
Weekly Retro (Friday): Sprint target vs actual, emails sent, leads sourced, revenue events, things shipped/broke, blockers hit.
Cron
- Monday 9 AM ET:
0 13 * * 1(leverage audit) - Friday 9 AM ET:
0 13 * * 5(retro)
Use OpenClaw cron (needs LLM access), not crontab.