Agent startup squad

Startup squad of agents

A startup that builds without measuring ships features into silence; one that debates without shipping never learns. A three-agent squad on a one-week clock forces the whole loop: build the smallest thing, put it in front of users, read the number, decide. The discipline is the cadence, not the headcount: a week is the unit, and every week ends with a keep, iterate, or kill call.

Team

• Builder (backend-engineer-role, frontend-engineer-role): ships the weekly increment.
• Marketer (developer-advocate-role): packages and promotes it on one channel.
• Analyst (data-scientist-role): reads the metric against the bet.

Shape: a fixed weekly loop, build to ship to measure to decide.

Method

1. Name one weekly bet. Pick a single metric to move (activation, week-one retention, signups) and the one thing you will ship to move it. Everything else waits for next week. Write it at the top of week.md.
2. Builder ships the smallest test of the bet. Not the full feature, the minimum that could move the number. The handoff is a running increment plus a changelog line.
3. Marketer packages one message. One channel, one audience, one call to action, drafted in launch.md. A human approves before anything public goes out.
4. Instrument before you ship, not after. The analyst confirms the metric is tracked and has a baseline; a number you start counting post-launch proves nothing.
5. Analyst reads the result against the bet. readout.md states the metric, the change, and whether the bet moved it, with the confounders named honestly.
6. Make the Friday call. Keep, iterate, or kill, recorded with a reason. The loop is a week, not "when it is done."
7. Enforce the kill rule. If the north-star metric has not moved across a set number of cycles (say four), stop the line of work rather than nursing it.

Run it

In Claude Code, run one loop per orchestrator pass: spawn the builder subagent, then the marketer (gated on human approval for anything public), then the analyst reading real metrics, each writing its file into a dated week directory the next reads. A /loop or scheduled run makes the cadence literal. Port it to CrewAI as a sequential crew re-kicked weekly, to AutoGen as a round-robin GroupChat, or to LangGraph as a cyclic graph with the decision node closing each week.

Signals it works

• Every week ends with a written keep, iterate, or kill, not a rolling backlog.
• The metric was instrumented before ship, so the readout is real.
• Killed bets actually stop; the squad is not carrying four half-live experiments.

Boundaries

This drives the operating loop, not the company: fundraising, hiring, legal, and real customer relationships are human work. Agents can draft marketing but a person owns the brand and approves public posts. The north-star metric and kill threshold are yours to set from real business context, not the squad's to invent.