Skip to content

CONVERSATION_LOG.md

Latest commit

 

History

History
98 lines (62 loc) · 5.54 KB

CONVERSATION_LOG.md

File metadata and controls

98 lines (62 loc) · 5.54 KB

Conversation Log — Building Simmer's Agent Swarm

This log documents the human-AI collaboration between Adrian (founder) and Claude Code that built Simmer's hackathon submission. Key moments from March 18-22, 2026.

March 18: "What if agents could run the company?"

Adrian: We have thousands of agents trading on Simmer. What if they could do more than trade — help run the platform itself?

Claude Code: Looked at Paperclip — an open-source control plane for AI-agent companies. It has org charts, goal hierarchies, task management, and audit trails. The key insight from our Wurk.fun experiment (EXP-B07): flat bounty boards produce spam. Structured roles + context produce quality.

Decision: Deploy Paperclip on Railway, create a "Simmer Labs" company with 0xSimmy as CEO.

March 18: Deployment Battles

Railway deployment was harder than expected:

  • Railway bans VOLUME directive in Dockerfiles
  • Embedded Postgres refuses to run as root (Railway's default)
  • Had to switch from PGlite to Railway Postgres
  • authenticated + private mode blocks unknown hostnames — needed allowlist
  • Dockerfile had wrong package name (@paperclip/ui → @paperclipai/ui)

Six hours of debugging later: Paperclip live on Railway with auth, org chart, and the first three tasks created.

March 19: The runId Problem

Adrian: Simmy can read tasks but can't write results back.

Claude Code: Investigated — Paperclip mutations require a runId that only exists when Paperclip invokes the agent. But Simmy polls on his own schedule via OpenClaw. Paperclip can't invoke Simmy because OpenClaw has no webhook endpoint.

Solution: Built an OpenClaw plugin (reactor pattern) that polls Paperclip and injects tasks into Simmy's prompt context. Write-back via board session auth. Also built the Task Bridge — a separate REST API for community agents.

March 19: Task Bridge Goes Live

Built and deployed the community task bridge (~200 lines of Express):

  • Agents authenticate with their Simmer API key
  • Bridge verifies against Simmer API, then proxies to Paperclip via board session
  • Community agents can only see tasks tagged "community"
  • Endpoints: GET /tasks, POST /tasks/:id/claim, POST /tasks/:id/submit

Deployed as separate Railway service. Registered 0xSimmy on Synthesis (ERC-8004 on Base Mainnet).

March 20: First External Agent Completes Tasks

The moment it became real: An external agent (Poly V2, agent ID 5d366a25) autonomously:

  1. Read the hackathon skill
  2. Registered on Simmer
  3. Browsed available tasks
  4. Claimed three tasks (SIM-19, SIM-20, SIM-21)
  5. Delivered high-quality work:
    • Competitive research on FelixCraft and Yoshi Zen with real URLs and revenue numbers
    • A specific product feature proposal (signal data backfill API)
    • An authentic 279-character tweet
  6. Submitted all three

All approved. 0.03 USDC paid on Base. Basescan TX.

No human intervention required beyond approval. The system works.

March 21: Maximizing Track Coverage

Adrian: We're only on one track. The builder guide says up to 10.

Claude Code: Mapped all 27 sponsor tracks against Simmer's capabilities. Found 9 that fit:

  • 5 require zero new code (just submission descriptions framing existing work)
  • MoonPay CLI Agents — integration already exists in the Simmer skill
  • Venice + Bankr — both OpenAI-compatible LLM gateways, config swaps only

The Venice angle: Route agent inference through TEE-secured models. Trading alpha stays private. Attestation-verified.

The Bankr angle: Self-funding agent loop. Agent earns from trading → pays for inference → makes better trades.

March 21: Final Push — 9 Tracks + AGENTS.md

Wrote AGENTS.md for agentic judges — health checks, curl commands, on-chain evidence, track-specific context for all 9 tracks. Expanded from 1 track to 9 via the Synthesis API. Updated submission with futarchy vision, community results, and integration documentation.

A second agent (clarence-live) submitted 3 more tasks — Spanish FAQ translation, meme concept, and a prediction market created via the API. All approved, 0.03 USDC paid on Base.

March 22: The Reframe — Open-Source Infrastructure

Adrian: Since Paperclip is open source, what if we released our task bridge as open-source middleware that any protocol could use?

Claude Code: Assessed the existing code — ~200 lines of Express, auth hardcoded to Simmer. To generalize: abstract auth (wallet-based EIP-191 for platform independence + configurable API key verification), add auto-payment on task approval, make everything config-driven.

Built the generalized Paperclip Task Bridge:

  • Pluggable auth: wallet signatures (any agent with a wallet) or API key verification against any platform
  • Auto USDC rewards on Base when tasks are approved
  • Repeatable tasks, structured discovery, agent-to-agent review
  • Deployed to production, replacing the Simmer-specific version

The submission reframe: "We built open-source middleware that turns Paperclip into an on-chain agent job board. Each protocol self-hosts their own — no shared platform, no vendor lock-in. simmer.markets is the first production deployment."

Also shipped a Jobs page at simmer.markets/jobs — a public job board showing available tasks with "Delegate to Agent" buttons.

Today → Now → Next: Agents trade on Simmer. They contribute to running it. They could govern it — through the very prediction markets they trade on.

This submission was partially built by the agent swarm it describes.