go-to-market-agent is a multi-agent engine that turns agentic web research and nine enrichment sources into a monthly competitive and market-intelligence report, with every section cleared by deterministic and LLM quality gates before it ships. The model interprets the data; it does not invent it.
You describe a brand and its competitor watchlist once in YAML. The engine researches the web, enriches competitors from up to nine data sources, writes facts-only sections, synthesizes a scoreboard and ICE-scored recommendations, and clears a two-stage quality gate before publishing to Markdown and JSON. It is the generalized form of a system run in production; the example tenant is a fictional US dog-food brand, Barkwell.
The demo runs the entire pipeline on bundled fixtures, with no API keys and no network.
git clone https://github.com/maxrihter/go-to-market-agent && cd go-to-market-agent
make install # uv sync --extra dev
make demo # full pipeline on fixtures; no keys, no networkIt writes a complete report to output/. The committed sample is in
docs/sample-report.md.
%%{init: {'theme':'base','themeVariables':{'background':'#faf9f5','primaryColor':'#ffffff','primaryTextColor':'#141413','primaryBorderColor':'#b0aea5','lineColor':'#b0aea5','fontFamily':'ui-sans-serif, system-ui, sans-serif'}}}%%
flowchart LR
cfg[/"tenant.yaml<br/>brand · niche · watchlist"/] --> research["Research<br/>Tavily supervisor"]
research --> enrich["Enrich<br/>9 data sources"]
enrich --> analysts["6 analysts<br/>facts-only sections"]
analysts --> synth["3 synthesizers<br/>scoreboard · summary · ICE recs"]
synth --> assemble["Assemble report"]
assemble --> gates{"Quality gates"}
gates -- fail --> stop1(["discard"])
gates -- pass --> review{"LLM reviewer"}
review -- reject --> stop2(["discard"])
review -- publish --> out["Report<br/>Markdown · JSON"]
classDef accent fill:#f3e2d9,stroke:#d97757,color:#141413;
classDef stop fill:#f4f3ee,stroke:#cfccc2,color:#87867f;
class out accent;
class stop1,stop2 stop;
A research supervisor delegates web research per section; six analysts write facts-only sections from the findings and the enriched competitor data; three synthesizers interpret them into a scoreboard, an executive summary, and ICE-scored recommendations; the report is assembled and then cleared by deterministic gates and an LLM reviewer before it renders.
- Data is sourced, not invented. Competitor metrics come from the enrichment sources, fabricated source URLs are dropped, and stale sizing years are rejected.
- The gate is two-stage. Deterministic checks (completeness, coverage, tautology, evidence chain, freshness, forbidden entities) plus an LLM reviewer decide what publishes; a rejected report is never persisted.
- The LLM layer degrades gracefully. Every role runs a primary, then fallback, then retry-with-hint chain, so one empty or malformed response does not fail the run.
The node-by-node design is in docs/ARCHITECTURE.md.
- Config-driven. Brand, niche, watchlist, and safety lists live in one
tenant.yaml; retargeting needs no code changes. - Multi-agent orchestration on LangGraph: a research subgraph, six analysts, three synthesizers, and a two-stage publish gate.
- Nine-source competitor enrichment: Instagram, Meta Ad Library, SimilarWeb, Wayback, Google Trends, YouTube, App Store, Ahrefs, and DataForSEO.
- Provider-agnostic LLM router (Anthropic, any OpenAI-compatible endpoint, Mistral, Google) with a primary, fallback, and retry path per role.
- Quality gates plus an LLM pre-publish reviewer that can reject a report, and an evaluation
harness (
gtm eval) that scores grounding, completeness, traceability, and clarity. - Plugin architecture for new sources, analysts, synthesizers, gates, outputs, and providers; month-over-month history in local SQLite; a hermetic demo that runs with no keys.
cp .env.example .env # add one LLM key, plus TAVILY_API_KEY and APIFY_TOKEN
gtm init # scaffolds config/tenant.yaml from the bundled example
$EDITOR config/tenant.yaml
gtm run --month 2026-05 # writes output/<report_id>.md and .jsonA live run needs one LLM key and a Tavily key for research; competitor enrichment uses an Apify token, and the paid SEO sources (Ahrefs, DataForSEO) are optional. Every config field is documented in docs/CONFIGURATION.md; the operational playbook is in docs/SETUP.md.
Each extension point is a small protocol plus a registry; ship a plugin or drop one in-tree.
Add a Source (a data connector), an Analyst (a report section), or a Synthesizer, Gate,
OutputAdapter, or LLMProvider. Run gtm plugins list to see what is registered. Details in
docs/EXTENDING.md.
| Layer | Technology |
|---|---|
| Orchestration | LangGraph (typed multi-node graph, subgraphs, checkpointing) |
| LLM access | Provider-agnostic router over LangChain provider SDKs, structured output |
| Data models | Pydantic v2 |
| Research and enrichment | Tavily, Apify, and per-source clients |
| Storage | SQLite by default, Postgres optional |
| CLI | Typer and Rich |
| Tooling | uv, ruff, mypy, pytest |
Full node-by-node design, resilience, and the gate model: docs/ARCHITECTURE.md.
- Report figures come from the enrichment sources; fabricated URLs are dropped and a report
reaches
output/only after clearing the gates and the LLM reviewer. - Secrets are read from
.envonly; nothing sensitive is committed and.env.exampleships empty placeholders. State is local SQLite by default. - Generalized from a private system with all client data, names, and strategy removed; the only brand present is the fictional Barkwell.
| Document | Contents |
|---|---|
| SETUP.md | Live-run playbook: keys, enrichment sources, scheduling |
| CONFIGURATION.md | Every tenant.yaml field |
| EXTENDING.md | Adding a source, analyst, gate, output, or provider |
| ARCHITECTURE.md | Node-by-node design, resilience, and the gate model |
Issues and pull requests are welcome; see CONTRIBUTING.md. Run
make lint && make test first (mocked, no keys required).