Angent

An autonomous deal-sourcing agent that finds, qualifies, and learns from startup outreach for angel investors.

Comment

Langfuse Trace

https://us.cloud.langfuse.com/project/cmqbhy5fc02lfad0jqk39tflt/traces

Inspiration

Solo angel investors miss great startups because sourcing is fragmented: launch sites, GitHub activity, thesis fit, outreach, and reply tracking all live in separate workflows. We built Angent to turn that process into one autonomous loop.

What It Does

Angent is a goal-driven deal-sourcing agent. Given a thesis, deadline, target reply rate, and email budget, it:

  • Finds startup candidates from public signals.
  • Scores them against the investor’s thesis.
  • Explains the fit.
  • Drafts personalized outreach.
  • Requires human approval before sending.
  • Tracks opens/replies.
  • Learns from outcomes.
  • Generates a markdown Deal Memo.
  • Optionally serves that memo behind an x402 payment gate.

Core metric:

$$ \text{reply rate} = \frac{\text{replies}}{\text{emails sent}} $$

How We Built It

The backend is a Python control loop with six stages: Planner, Scanner, Qualifier, Writer, Governance Gate, and Optimizer.

We used:

  • ClickHouse as the shared blackboard for companies, emails, outcomes, loop state, publications, and paid fetches.
  • TrueFoundry/OpenAI-compatible inference for explanations and outreach drafts.
  • Pioneer as an optional adaptive scorer with heuristic fallback.
  • Airbyte/GitHub, Hacker News, and Hugging Face as signal sources.
  • OpenUI for generated investor-facing surfaces.
  • x402 to gate paid access to generated Deal Memos.
  • Langfuse-style tracing and stage logs for observability.

Run command:

python -m angent.main --demo

What We Learned

The hard part of agent autonomy is not just generation. It is state, control, and recovery. Angent needed durable memory, explicit goals, budget limits, approval gates, fallback scorers, and clear per-stage logs so the agent could keep moving without becoming unsafe or opaque.

Challenges

The main challenge was making many integrations behave like one system. We had to handle unavailable APIs, partial data, slow inference, failed persistence, and email safety. We also separated two output paths clearly: OpenUI generates interactive human-facing views, while the Publisher emits markdown Deal Memos for cited.md/x402 distribution.

Outcome

Angent is a working prototype of an autonomous investment analyst: it sources companies, qualifies them, drafts outreach, records outcomes, and improves the next loop from real feedback.

Built With

Share this project:

Updates