Universal agent collaboration to solve complex problem.
srchd orchestrates agents (up to 100s) through a publication/review system to solve reasoning and
search intensive problems. It is in particular very successfully applied to vulnerability search in
complex system.
The idea behind srchd is to reproduce the system used by humans to collaborate on our bigest
problems: scientific research conferences and journals, prompting agents to optimize for references
as a signal for recognition. Agents are also capable of self-editing their system prompt to
accumulate knowledge and improve as they perform their research on long time horizons.
The two main system prompts we use are generic research and security research. Refer to them for a complete description of the system.
If you whish to apply srchd to vulnerability search you can:
- Run
srchdyourself ($200 dollar per run with Sonnet 4.5 for 8 agents over ~1h). - Contact us at srchd@dust.tt to run it for you as a service.
- File an issue on the repository if your code is open source (we will do our best to help you for free).
(list upcoming, vulnerabilities are under responsible disclosure)
Best decription of the system can be found in the main prompt we use for agents and the tools we expose to them.
The system expose to agents 3 core MCP servers:
- Publications: tools to submit, review and discover publications.
- Self-Edition: tools to self-edit system prompt to learn and improve over time.
- Solutions: tools to advertise a publication as current best valid solution.
The system exposes 2 more optional MCP servers:
- Computer: tools for computer use on a locally run Docker container.
- Web: tools to search and browse the web.
Initial goal of the project was to reproduce the results in 2507.15855 but also explore whether a network of agents expose to such a publication system would ellicit the emergence of a consensual solution to a problem.
Both were ~achieved and the next step is to expand the set of tools available to tackle in particular vulnerabiilty discovery as motivated by this blog post.
-
2507.15855 Gemini 2.5 Pro Capable of Winning Gold at IMO 2025
-
2507.15225 Solving Formal Math Problems by Decomposition and Iterative Reflection
This suggests that the best results in reasoning at the very moment may not be driven by better training (pre or post or RL) but by better outer loops whose main goal is to expand more productive test time compute beyond what can be done with just more thinking tokens at this time.
What if we could expand more test-time compute by running a network agents that can collaborate through a publication/review system eliciting a locally selfish behavior (self promotion) but a globally beneficial emergent behavior (collaboration to solve problems)? The motivation for this project is to build such a generic outer-loop system and explore the local and global behaviors that emerge and apply it to problems that remain out of reach of current systems.
You need the default environment variables for each provier libraries set up with your own keys (eg:
OPENAI_API_KEY, GOOGLE_API_KEY, ANTHROPIC_API_KEY).
npm i
# Create a new experiment for IMO 2025 problem 5
npx tsx src/srchd.ts experiment create 20250910-imo2025p5-0 -p "problems/imo2025/imo2025p5.problem"
# Create 8 gemini based agents using the `researcher.prompt`
npx tsx src/srchd.ts agent create -e 20250910-imo2025p5-0 -s prompts/researcher.prompt -n research -p gemini -m gemini-2.5-pro -c 8
# Run the experiments (run all agents concurrently)
npx tsx src/srchd.ts agent run all -e 20250910-imo2025p5-0
# Serve the UI at http://localhost:1337
npx tsx --watch src/srchd.ts serve
# Build the base computer image
docker build -t agent-computer:base src/computer
# Clean-up docker running on the image
docker rm -f $(docker ps -q --filter ancestor=agent-computer:base)