More experiments than you can run by hand
Parallel subagents explore many directions at once, each in its own isolated git worktree. You stop rationing experiments to whatever fits in a sprint.
Autoresearch
Autoresearch, as infrastructure for your stack
Your team already runs autoresearch by hand — proposing a harness tweak, running the eval, reading the scoreboard, keeping what wins. EVO turns that loop into infrastructure: agents propose changes, prove which ones beat your baseline, and ship only what's measurably better — across code, agents, models, configs, and prompts. Continuously, with a receipt for every change.
What autoresearch gives your team
If you own an eval suite and hand-tune a complex harness, autoresearch is the difference between iterating on intuition and iterating on evidence. EVO runs the experiments you don't have time to run — in parallel, gated by your safety checks — and only the changes that move your metric survive.
Every win is gated and provable
A change only lands if it beats your baseline and clears your regression tests and invariants. A faster wrong answer never wins.
A receipt for every change
Each kept improvement comes with the experiment, the score delta, and the discarded alternatives — so progress is auditable, not a black box.
How EVO runs autoresearch
The same loop you run manually — automated, parallelized, and safety-gated.
- 1
Define what "better" means
Pick the metric to move and the direction — higher accuracy, lower latency, cheaper tokens. A benchmark turns that into a number the loop can chase.
- 2
Propose a change
An agent reads the codebase, prior traces, and discarded ideas, then forms a hypothesis and makes one concrete, isolated change to test.
- 3
Run the experiment
The change runs in its own sandbox or worktree against the benchmark, scored independently so attempts never collide or corrupt your main branch.
- 4
Gate for safety
Regression tests and invariants run as gates. An experiment that breaks something is discarded — even if its score went up.
- 5
Keep or discard
Only changes that beat the baseline and clear every gate survive. The rest are thrown away, with the reasoning recorded.
- 6
Raise the baseline and repeat
A kept win becomes the new baseline, and the loop runs again — exploring multiple directions and compounding verified progress.
Hand-tuned harness vs. autoresearch as infrastructure
| Criterion | Your DIY harness | EVO |
|---|---|---|
| Throughput | One experiment at a time, between other work | Parallel subagents, each in its own isolated git worktree |
| Search strategy | Greedy hill-climb on whatever you tried last | Tree search — multiple directions fork from any committed win |
| Safety | You remember to run the regression suite — usually | Gates discard any change that breaks a test, even if the score rises |
| Shared learning | Each attempt and dead-end lives in someone's head | Failure traces and discarded hypotheses are shared across agents |
| Scope | You tune the one layer you have time for | Compounds across code, agents, models, configs, prompts, and A/B tests |
| Setup | Hand-build and babysit your own harness | Discover explores the repo and instruments the benchmark for you |
Further reading
Explainer
What is autoresearch?
The category, defined: an autonomous loop that proposes changes, scores them against a baseline, and keeps only what measurably wins.
Claude CodeAutoresearch on Claude Code
EVO instruments your existing Claude Code prompts, tools, and subagent graphs as the search space — and proves which changes beat baseline.
Ship improvements.
Not guesses.
Join teams turning every change into
measurable, verified progress.