GTM Agent Arena
A multi-stage AI system that turns messy startup context into a structured go-to-market strategy.
A Context Compressor, then ICP, Pain, Channel, and Offer agents, each answering one question and building on the last. A Critic reviews the whole chain for contradictions and overconfidence before a Final Report reconciles what survives into a scorecard, three experiments, and a 7-day plan.
Watch it, then run it.
GTM Agent Arena passes a startup through specialist agents for customer selection, pain, channels, positioning, criticism, and final synthesis. The demo below uses the Canadian Shawarma Report as the test case.
Run the real seven-stage pipeline → in your browser, with your own Anthropic or OpenAI key. The Shawarma context is preloaded; nothing is stored.
Why multiple agents?
The useful part isn't making several models talk. It's forcing an ambiguous GTM problem through distinct decisions, then giving a separate critic permission to challenge the conclusions before synthesis.
The pipeline.
Each agent has one job and a fixed JSON contract. Later agents receive the compressed context plus every prior agent's structured output — never the raw founder text.
- Context
Compresses raw input into one brief: category, current audience, current channels, and the founder's own stated uncertainty.
- ICP
Scores candidate segments on pain intensity, willingness to pay, ease of reach, and speed to first sale, then names a launch wedge and one segment to avoid for now.
- Pain
Ranks candidate pains by urgency and current workaround, and picks the one worth leading with.
- Channel
Scores channel fit and names both where to start and where explicitly not to start yet.
- Offer
Diagnoses the current positioning, then rewrites the headline, subheadline, CTA, and objection responses.
- Critic
Reviews every prior output for contradictions, the weakest assumption, overconfident claims, and missing data — then revises the strategy.
- Final Report
Synthesizes the surviving recommendations into a scorecard, three GTM experiments, and a 7-day action plan.
Under the hood.
- App
Next.js App Router and TypeScript; a dashboard, a report form, and a report-detail view with agent outputs and outreach tracking.
- Models
The six structured agents run on a smaller model (
gpt-5-mini) in JSON mode; the Final Report runs on a stronger model (gpt-5) for the prose synthesis. - Reliability
Each model call retries up to three times with backoff, and a regex fallback recovers JSON the model wrapped in prose or a code fence.
- Data
Reports persist to Supabase Postgres when configured, or a local JSON file store otherwise — the same report shape either way.
Outcome tracking.
A generated report isn't the end of the loop. Once a report is actually sent to a founder, the dashboard records whether they replied, a 1–5 usefulness score, a quote, and free-form notes — and rolls those up into a response rate and an average usefulness score across every report sent.
Sending, replying, and scoring are recorded by hand, not inferred. The loop exists to find out which recommendations were actually useful, not to simulate that they were.
Status.
The production app is a private Next.js service run against the OpenAI API, with a dashboard and outcome tracking — an internal tool, not a consumer product. The bring-your-own-key runner on this site is a static, self-contained build of the same seven prompts and stage order, with no accounts, no payments, and no persistence.