Start — idea in, launch packet out

An agent that takes a startup idea and returns what the founder has to do next, unattended: a staged roadmap, a compliance checklist, first hires, funding targets, drafted documents, and a queue of only the decisions the CEO must make. Fallback entry for the "Build an Agent" Hackathon, Corgi Cafe, San Francisco, Oct 7, 2026.

Status (7 Oct 2026, 3:35 PM): works end to end on our machine. Run today: Agent37 research (4 turns), OpenAI structured packets (5), Supabase feed read and written by worker and page, Monid live search (15 results in one run), the decision queue, and one steer rework of a packet after a saved decision. An idea typed into the page was picked up by the watching worker and came back as a packet in 325 s with no terminal step. Not done: hosting (the page runs locally), any measurement of the manual workflow, and anything that files, pays or sends. 43 unit tests pass; the network clients are covered by the live runs, not by unit tests.

What it does

Idea in -> packet out, no human step until the decisions:

1. Research: an Agent37 instance works through what this company has to do.
2. Structure: OpenAI structured output turns the notes into one packet JSON.
3. Audit, in code: a "sourced" claim whose URL is missing or does not resolve is downgraded to "unverified". Tasks per owner (agent / CEO / professional) are counted, not estimated.
4. Store: the packet is written to Supabase; the page reads it as a feed.
5. Steer: the CEO answers the decision queue on the page and presses "Rework the plan"; the agent rebuilds the packet around those choices and publishes a revision.
6. (Before research) Monid live web search gathers leads on regulations, investors, competitors.

It drafts and lists. It does not file, pay, send, sign, or give legal advice.

Stack

• Agent37 Cloud APIs: the unattended research worker (required by the hackathon)
• OpenAI: structured output for the packet
• Supabase: stores packets, decisions and requests; the web page reads and writes the feed
• Monid: live web search (regulations, investors, competitors) handed to the agent as leads

Layout

• agent/packet.py packet schema, validation and the deterministic audit
• agent/clients.py clients for Agent37, OpenAI, Supabase and the Monid CLI, plus the link check
• agent/worker.py the pipeline: one idea, --steer ID, or --watch to answer the web page
• agent/provision.py creates the Agent37 instance: python3 -m agent.provision
• web/index.html the feed page
• supabase/schema.sql a dedicated table, for later; today rows go to the existing papers table
• docs/SCOPE.md the vision, today's slice, existing players, decisions, demo script
• docs/REQUIREMENTS.md hackathon rules, sponsor plan, submission checklist
• docs/SUBMISSION.md paste-ready answers for the submission form

Run

cp .env.example .env                 # fill it in; never commit .env
python3 -m pytest agent -q
python3 -m agent.provision           # once: creates the Agent37 instance (billable, per minute)
python3 -m agent.worker "your idea"  # one packet from the terminal
python3 -m agent.worker --steer 7    # rework packet 7 around the decisions saved on the page

# No-terminal mode: leave the agent watching, do everything on the page
python3 -m agent.worker --watch
python3 -m http.server 8737 --bind 127.0.0.1 --directory web   # then open http://127.0.0.1:8737

web/config.js (Supabase URL and anon key for the page) is written by the worker and is not committed. Optional live search: npm install --prefix .tools @monid-ai/cli, then .tools/node_modules/.bin/monid keys add --key <key> --label <name>.