Loading page…
Loading page…
Darwin Compiler studies how agent definitions become executable operations and compatible interfaces, with their meaning intact.

An idea, an existing API, and a working agent are different starting points. Each must resolve to an explicit account of what the agent can do: named operations, input and output contracts, connections, permissions, approval rules, and limits.
Agent Builder is the authoring experience, supported by the Darwin Builder AI coding system. Generation can help refine a definition; compilation checks and translates the definition into a form that a target can execute or expose. Unresolved assumptions must remain visible.
An operation can call an existing endpoint, use a versioned node, or compose a workflow of nodes. A custom implementation also needs its source, entry point, dependencies, and execution requirements. A Python file alone does not specify a reproducible runtime.
We study how to validate the connections between these contracts and produce versioned artifacts. An existing service keeps its implementation on its host; its adapter must preserve authentication, inputs, outputs, and error behavior.
| Source definition | Compilation question |
|---|---|
| Existing endpoint | Can its request, response, authentication, and failure contracts be mapped without changing behavior? |
| Node or workflow | Are pinned nodes compatible, are required inputs available, and do permissions constrain every step? |
| Custom code | Can the declared source and dependencies build reproducibly in the required execution environment? |
A person may need an accessible form, progress view, and approval screen. An agent needs structured schemas, action identifiers, and explicit state. The presentation differs; the underlying operation and authority must agree.
Compiler research asks which parts of that contract can be preserved across MCP, API, and other supported protocol adapters. Composer adapts the presentation to the caller, host capabilities, and current operation state. A host that cannot render components still needs a useful structured or text representation.
Existing component libraries can supply the human view and agent presentation contracts. Missing states, unsupported host capabilities, and incomplete adapters must be reported rather than silently replaced with a promise of compatibility.
A valid schema is one check. Equivalent behavior across targets is another. Our research questions include contract checking, workflow compatibility, permission preservation, and differential tests of the same operation through different interfaces.
Evaluate successful results, validation errors, approval waits, cancellation, and failures. Pin evidence to the operation, implementation, and adapter revisions; a change should invalidate the evidence it affects.
A Darwin Agent ZIP is the portable source package. Destination exports are derived artifacts. Building either package does not establish that its server is live, its client interaction works, or a directory has accepted it.
The compiler prepares and validates an execution artifact. Runtime accepts a submitted invocation, creates its job, and manages tasks, approvals, recovery, and results. Opening an operation interface should not start the work.
Research here concerns translation and behavioral guarantees. Hosting, dependency installation, and archive generation support the system; they do not by themselves demonstrate those guarantees. Agent Builder exposes the authoring process, while publication and live client testing remain separate stages.
Explore how Runtime coordinates authorized work, tracks progress, and handles recovery.
Explore Darwin Runtime