Loading page…
Loading page…
Darwin Runtime studies how agents coordinate work, preserve intent, and recover without losing control of the outcome.
Real tasks rarely end with one response. They move through inputs, discovery, decisions, provider work, and results. The runtime research question is how to carry the original goal through that entire sequence.
A sequential workflow waits for one step before starting the next. Some work really has that dependency; other work can proceed independently. Our research asks how agents can coordinate both without losing task ownership, access boundaries, or evidence.
Each result supplies the next step’s input.
Separate evidence can be gathered before a shared decision.
Delegation needs more than an agent name. The receiving participant needs the relevant objective, validated inputs, permitted effects, and a clear definition of the expected result. Shared progress must make blocked work and missing decisions visible.
We study how to carry those facts across dependent and parallel tasks while keeping sensitive context and approvals scoped to the work that requires them.
Acceptance, provider work, and verification are different events. A useful execution record preserves that distinction so a reconnection can reconcile state instead of starting the action again.
{"stage":"request","objective":"<original task>","episode":"<durable id>"}{"stage":"approval","state":"<required decision>"}{"stage":"provider","reference":"<execution receipt>"}{"stage":"verification","evidence":"<independent outcome check>"}The execution benchmark evaluates the complete system—from the user’s request to an independently verified outcome. It complements discovery ranking and individual-agent reliability without blending their scores.
| Measure | Boundary |
|---|---|
| Verified completion | Stop on the required outcome or deadline, not the first response or thread creation. |
| Elapsed time | Include discovery, setup, approvals, provider work, failed attempts, and recovery. |
| Total cost | Account for metered model, tool, search, and provider fees. Mark unobservable costs explicitly. |
| Human intervention | Use a frozen approval and clarification policy; record substantive assistance. |
A disconnect after possible acceptance is not permission to repeat an action. Recovery should reconcile provider state, retain durable identifiers, and respect cancellation and approval boundaries.
The expected answer and the scoring oracle belong outside the production worker. Workflow completion, internal processing, and original-objective success need separate counts.
Explore how independent verification separates an agent’s claim from a checked outcome.
Explore Darwin Verifier
