Skip to content

Agentic Runtime

The control layer that keeps AI from guessing.

The runtime executes a workflow step by step and validates each stage before the next one runs. If information is missing, a requirement is not met, or a quality check fails, the run stops instead of silently continuing — and every run produces a traceable result. This is what turns AI generation into controlled, auditable execution.

What a runtime is for: taking a process off the page and running it for real, one step at a time, checking each step before the next and keeping a record of what happened. Without it a workflow is a diagram; with it, it is a run someone can inspect afterwards.

What it does

  • Runs the process exactly as published. Every step delegates to a specialized agent taken from a versioned catalog, so the procedure on paper and the work actually performed cannot drift apart.
  • Checks each step before the next one starts. Eval-gates — checkpoints that inspect a step's output before the following step is allowed to start — decide whether the chain proceeds; a failed blocking check stops the run instead of passing the problem downstream.
  • Sends an incoming request to the right process, or says why it cannot. A dispatcher (WF-000) reads the free-form brief and answers NO_MATCH when no process fits, PARAMS_MISSING when the request is too thin to act on — rather than guessing. That check covers all ten workflows since v0.10.0.
  • Never writes to the catalog. It only reads from it, through pinned, versioned imports, so running a workflow cannot quietly change the assets it ran on.

Highlights

  • How it fits together at a glance — the two entry modes, the gates, and the traces — see Architecture.
  • Proven live, end to end — see Live proofs.
  • Routes the request, and refuses it, before any work starts — see the dispatcher runs.
  • ISO-aligned quality and licence audits — see Audits & licences.
  • Clean, dependency-light core (@anthropic-ai/claude-agent-sdk + ajv).

▫ claude-agentic-runtime v0.18.0 — a headless orchestration engine for multi-agent workflows: it has no user interface of its own, and is driven by the workflow files themselves.

Source repository →