YOUR REPOSITORY. A RESEARCH LAB.
Turn “make this better” into a measured research loop. Test changes in isolated worktrees. Keep progress that earns an independent review.
Local-first. Copilot-native. Nothing merges without you.
Branch in parallel. Extend promising experiments. Recombine ideas and measure again.
Scroll the graph to follow every branch
Base: exp_0004 · Donor: exp_0005
Keep the base's improvement and transfer a compatible idea from the donor. Remeasure the combination, pass the gates, and earn a new independent review. Nothing merges into main.
Select a node to trace its parents and donors. Illustrative run, not benchmark results.
A research loop.
Not another hosted platform.
THE RESEARCH LOOP
Optimization needs a trustworthy measuring stick. Olo separates building that measurement from changing the code it measures.
/olo-exploreUnderstand the repository. Choose a goal with real headroom. Construct or adapt a benchmark and regression gates inside the baseline worktree.
YOU GET A checked baseline. Frozen measurement. A clear starting point.
/olo-optimizeRead failure traces. Explore different hypotheses in isolated worktrees. Measure each candidate against the same benchmark, then review the evidence.
YOU GET Reviewed improvements. Useful alternatives. Lessons for the next round.
The last step stays human. Inspect the winning branch. Decide what to merge.
Meet the six agent rolesMORE THAN A SCORE
A promising number is not enough. Olo keeps the source, task-level evidence, and review decision connected.
Source-bound checks establish repeatability before the baseline. Protected files keep the measuring stick fixed during experiments.
check → assess → baseline
Measured candidates wait for independent approval. Numeric gain floors and critical-task limits separate progress from noise.
pending-review → approved
Retain task specialists. Revisit approved ancestors. Combine compatible ideas with their donors and contributions recorded.
frontier · retain · recombine
Recorded probes and evidence-linked lessons make failed directions useful to the next experiment instead of lost in a chat.
probe · traces · learn
Bound the evaluations. Respect ceilings and stalls. Run a configured final test once, then close that measurement version.
bound → select → finalize
A local dashboard connects the experiment tree to gates, task deltas, diffs, and logs. Read or download the generated run summary.
python olo.py report
A perfect tiny benchmark is a ceiling, not production proof. Olo makes that boundary explicit.
FROM REPO TO FIRST EXPERIMENT
Try the included five-case policy router in a fresh clone. It is intentionally small enough to understand, including the cases it gets wrong.
No package install for Olo's controller. Copilot usage and your target's dependencies are separate.
Bring Olo to your own repositorygit clone https://github.com/abpatramsft/olo-autoresearch.git
cd olo-autoresearch
python olo.py --help
copilot --agent=olo-explorerUse your Python 3 launcher if it is named python3.
Use /olo-explore on examples/tiny-policy-agent/agent.py. Use the existing benchmark and gate, protect both measurement files, and prepare a checked, independently approved baseline. Stop at phase=ready-to-optimize. Do not merge.
A FEW USEFUL BOUNDARIES
No. Olo's controller, worktrees, dashboard, and saved evidence live in your repository on your machine. This website is a static introduction, not the dashboard. Copilot itself uses its configured model services; your benchmarks may also use external services.
Experiment edits happen in allocated Git worktrees. Benchmark construction belongs in the baseline worktree. A winning experiment stays on its own branch until you explicitly decide to merge it.
Olo is loosely inspired by Evo and its experiment-driven approach. It is independently written for GitHub Copilot CLI, with a particular focus on checked baselines, binding review, traceable research, and final-test discipline. It is not an Evo plugin, and this is not a feature-parity comparison.
Olo is a local prototype, not a security sandbox or proof of production reliability. Worktrees do not isolate hardware or network access. Independent review is a workflow requirement, not identity authentication. Use representative benchmarks and fresh final questions.
LESS GUESSWORK. MORE EVIDENCE.