Coding agent
Claude
Codex
Cursor
/weco improve validation accuracy
Goal set
Two agents. One loop.
Your coding agent hands Weco a goal and an eval. Weco searches a tree of code changes, returns the best patch with evidence, and lets you steer the next branch in chat whenever you want.
we·co
tree searchA search tree, not a black box.
Powered by AIDE, our open-source tree-search engine. Every attempt is a node - Weco keeps the winners, branches further, and shows the whole lineage, so you can follow, redirect, and bound the search.
best-solution.py
Steerability
guide itFork promising subtrees and add constraints in natural language to redirect the search whenever you want.
Trust
bound itBudget-bounded by design. It won't break your pipeline halfway or run forever - you set the limits up front.
Observability
see itEvery attempt is a node - inspect its diff, metric, and logs, and follow the whole lineage of how the search got there.
Proven on real problems.
Fraud Detection
AUC 0.91 → 0.93
for end-to-end transaction-fraud ML pipelines
Prompt Engineering
+32.4% accuracy
for vision-language chart-to-CSV extraction
Inference Optimization
+47.7% speedup
for causal self-attention in GPT-style LLMs
Hyperparameter Optimization
−91.3% RMSLE
for molecular property prediction in materials science
Don't take our word for it.
So amazing to see something built by this team that's substantially underpinning and influencing OpenAI's agentic roadmap.
Edward Grefenstette
Director of ResearchGoogle DeepMind
Weco consistently surfaced improvements across my entire ML pipeline that I would never have discovered manually. It doesn't just accelerate development - it enables a fundamentally different way of building intelligent systems.
Christian Schwarz
Head of AIStandard CharteredPortfolio optimization
OpenAI is nothing without AIDE. Really cool to see @WecoAI's agent framework give o1 such a huge uplift.
Minqi Jiang
AI Researcher
Weco is how we go from 90% to 99%. We use it as a second phase after training - and the improvements it finds have directly helped us close deals.
Jun Cho
ML EngineerVardera LabsCounterfeit detection & authentication
Setting up Weco was incredibly easy; it instantly understood what to do, and I barely had to touch it.
Dan Schwartz
Founder, ArbiumOS Context Layer
We used Weco to optimize our voice AI prompts - accuracy jumped from 82% to 96%, with entire failure categories going from 0% to 100% in just a few iterations.
Tyler McGraw
ML Engineer, VeritusVoice AI prompt optimization
We used Weco on a bioacoustics research problem where we had to design new algorithms from scratch. It replaced weeks of manual heuristic work.
Param Mohapatra
Applied Math PhDCornell UniversityJennifer Sun's Lab
Flexible pricing
Try it on a real run
Free
- 20 free credits ≈ 100 steps of Weco-hosted autoresearch.
- Free access to Weco Observe: bring your own autoresearch agent and use the Weco dashboard for observability.
Bring your own keys
BYOK
- Use your OpenAI, Anthropic, or Gemini keys.
- Up to 10k experiments per month.
Managed inference at scale
Pay as you go
- Everything in Free, plus:
- Access to all models.
- Priority support.
Frequently Asked Questions
Turn your codebase into a self-improving system
Connect your coding agent to the dashboard and turn any eval into a self-improving loop.