Autoresearch on the workflows you already run
EVO instruments your existing Claude Code agents, prompts, and subagent graphs as the search space — no rewrite, no separate harness. The workflow you ship is the thing being optimized.
EVO × Claude Code
EVO ports autoresearch onto Claude Code dynamic workflows
If your team builds dynamic agent workflows on Claude Code, you already tune them by hand — rewriting prompts, reordering subagents, swapping tools, then eyeballing whether it got better. EVO turns that into autoresearch: agents propose changes to your Claude Code workflows, prove which ones beat your baseline, and keep only what's measurably better — gated by your safety checks, with a receipt for every change.
What EVO gives your Claude Code workflows
Claude Code makes it easy to build dynamic, multi-step agent workflows. It does not tell you which version is actually better. EVO runs that experiment for you — porting autoresearch onto the workflows you already ship — so prompt, tool, and orchestration changes compete on evidence instead of intuition.
Only measurable wins land
A proposed change to a prompt or orchestration step only survives if it beats your baseline on your eval and clears your regression checks. A faster wrong answer never ships.
A receipt for every change
Each kept improvement comes with the experiment, the score delta, and the discarded alternatives — so you can hand a prospect a clear before/after, not a vibe.
How EVO runs autoresearch on Claude Code
The same loop your team runs by hand on a Claude Code workflow — automated, parallelized, and safety-gated.
- 1
Point EVO at your workflow
Connect the Claude Code workflow you want to improve and the eval that defines "better" — task success, latency, or token cost. EVO maps prompts, tools, and subagent steps as the search space.
- 2
Propose a change
An agent reads the workflow, prior runs, and discarded ideas, then makes one concrete change — a tightened prompt, a reordered subagent, a swapped tool — to test against the baseline.
- 3
Run it in parallel
Each candidate runs in its own isolated git worktree against your eval, scored independently so dynamic workflows never collide or corrupt your main branch.
- 4
Gate for safety
Your regression tests and invariants run as gates. A change that breaks a tool call or a guardrail 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 for the next round.
- 6
Raise the baseline and repeat
A kept win becomes the new baseline, and the loop forks again — compounding verified improvements across your Claude Code workflows continuously.
Hand-tuning a Claude Code workflow vs. autoresearch on it
| Criterion | Hand-tuning in Claude Code | EVO autoresearch |
|---|---|---|
| How changes are tested | Edit a prompt or step, rerun, eyeball the difference | Every change runs against your eval and is scored against the baseline |
| Throughput | One workflow variant at a time, between other work | Parallel subagents, each in its own isolated git worktree |
| Search strategy | Greedy tweaks on whatever you tried last | Tree search — multiple directions fork from any committed win |
| Safety | You remember to recheck guardrails — usually | Gates discard any change that breaks a test, even if the score rises |
| Scope | You tune the one prompt or step you have time for | Compounds across prompts, tools, subagent graphs, and configs |
| Evidence to share | "It feels better" — no audit trail | A receipt per change: experiment, score delta, discarded alternatives |
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.
ProductAutoresearch as infrastructure
For teams who own an eval suite and hand-tune a complex harness. EVO runs the experiments in parallel and ships only measurable wins.
Ship improvements.
Not guesses.
Join teams turning every change into
measurable, verified progress.