cekura-self-improving-agent

Provide the production call ID or Cekura result ID and the skill fetches the failure, reproduces it in simulation with a must-fail gate, then iterates prompt or tool-config edits on a non-production clone until the failure set passes verification thresholds.

Verification re-runs Cekura simulation scenarios against the edited agent. Deterministic bugs need 2/2 passing runs with the trigger active; stochastic bugs need at least 80% of 5-10 runs passing, followed by a full-set regression sweep.

Yes, provider-dashboard agents on VAPI, Retell, ElevenLabs, and Bland are supported via pre-filled manifest recipes, including cloning the provider agent and tools so iteration never touches production. Custom stacks with config in a repo, database, or prompt registry are also supported.

It is a per-project file declaring where the agent's config lives and the commands to read, render, apply, deploy, and verify it. The skill treats it as untrusted infrastructure code: commands are registered verbatim, parameters are typed and escaped, and production environments are refused outside the Promote phase.

The must-fail-first invariant requires a recorded failing Cekura simulation result (repro.json) before any edit is proposed, preventing speculative prompt changes. A failing unit test or production log alone never satisfies this gate; only an explicit user override can bypass it, and the output is then marked as an unverified hypothesis.

No. All iteration happens against a non-production environment or cloned agent. Production changes occur only in the explicit Promote phase, which requires user confirmation, a rendered diff, and a declared rollback path per component.