October 2026
shipped

Evidence-first Agent Supervisor

An agent needs to remember what happened, especially when its tools stop answering.

Multi-step agent workflows combine research, external actions and human review. A timeout can leave a message sent but unconfirmed, while a saved form or a promise can be mistaken for a completed outcome. I built a supervisor playbook to keep those states distinct and make the next run resume from evidence instead of repeating uncertain actions.

  • Agent Skills
  • Markdown playbooks
  • Evidence ledgers
  • Capability adapters
01

Design boundaries

Account ownership

Basis — Architectural rule in the published skill, not a performance measurement
Single writer

Final submission

Basis — Explicit boundary in the published workflow, not an automated completion claim
Human review

The project

I built an operational supervisor for coordinating discovery, verification, authorized outreach, document preparation and review queues. The public release extracts its reusable coordination pattern into an installable skill. Career operations provide the example domain; this case study is about the system’s architecture and failure handling.

Source and installation · Design story · Synthetic demonstrations

Evidence before state transitions

The supervisor separates discovery from official verification, a referral promise from a received referral, and a portal draft from a confirmed submission. An uncertain send stays uncertain until conversation history resolves it. An empty new queue does not erase historical actions.

One supervisor owns accounts and shared writes. Optional research workers return compact evidence or prepare materials in isolated directories. On a host without subagents, those roles run sequentially with the same ownership boundary.

Designing for interruption

Incremental checkpoints record completed items, partial source coverage, uncertain actions and the next cursor. After a model limit or tool failure, the next run reconciles evidence before retrying. Missing inbox access produces a blocked check, not an invented claim that nobody replied.

Connectors, browsers, schedulers and model selection are capability adapters. The instructions require no particular model or provider. Claude Code, Codex, Cursor and other compatible agents can load the skill; available account tools and permissions still determine what they can execute.

Iteration and validation

The underlying guidance evolved to prioritize unfinished preparation, separate referral stages and maintain evidence-based review queues. The public version removes personal configuration and uses fictional fixtures. It preserves final employer submission as a human action.

The release includes metadata and packaging checks plus realistic synthetic scenarios. These are validation artifacts, not a controlled experiment. I am not claiming measured efficiency gains, conversion improvements or a per-run subscription cost.

Trade-off

Keeping one supervisor responsible for account operations limits parallelism, but gives every external action and shared-state update a clear owner. Instructions also rely on the host following them; this package is not an enforcement engine or a hosted service.

◆ What I would build differently

I would add repeatable offline evaluations for individual agent hosts and measure representative runs where usage metadata is available. The current release documents scenarios and boundaries; it does not claim benchmark gains, token savings or universal integration compatibility.