Coding agents need a delivery system

From May to June 2026, I contributed to affiliate-mcp with my collaborator and a few coding agents, mainly Claude Code and Codex. My work covered affiliate-network integrations, data pipelines, technical architecture, and agent-to-agent delivery collaboration.

Project note: This article describes that collaboration as it operated during that period. I am no longer an active contributor to the repository.

The agents helped us write code faster. Coordination soon became the harder part: deciding what should ship, keeping parallel work coherent, reviewing changes, and knowing when a human had to make the call.

The outcome became the unit of work

A prompt and a branch work well for an isolated fix. Parallel changes need a clearer boundary. We define each workstream around one outcome, its owning layer, dependencies, failure modes, and excluded scope. The same thread inspects the repository, implements the change, validates it, reviews the diff, and prepares the handoff.

I kept a persistent Chief of Staff thread above those workstreams. It tracked active work, queued ideas, dependencies, reviews, and decisions that needed me or my collaborator. The workstream threads handled patches. The Chief of Staff kept the project map current.

Repository skills became operating procedures

We moved delivery rules into the repository. AGENTS.md explains the architecture, authoritative files, reading order, and decision boundaries. delivery-steward guides a bounded change from inspection to proof. prepare-for-review defines the evidence a pull request needs. review-pr checks contracts, correctness, tests, architecture, customer clarity, and documentation.

These files are version-controlled operating procedures. When a failure repeats, we improve the procedure once. Every future agent receives the same guidance. Git history also shows how the collaboration rules evolved.

Human authority should follow risk

Agents can make routine, isolated choices after reading the owning files and existing patterns. Public contracts, authentication, security, write actions, payments, deployment architecture, and cross-client behaviour receive human review.

During the collaboration, technical architecture and product trade-offs were my focus. Affiliate-domain truth and customer judgement stayed with my collaborator. For consequential questions, the agent recommended a direction, explained the trade-offs, and escalated the decision. This gave agents room to work while preserving a clear audit trail.

Reviewer agents can repair scoped issues

A reviewer agent can often repair a small, clear defect itself. It inspects the branch, makes the smallest correction, runs focused checks, and reports what changed. This removes a slow handoff through a human.

The correction must be scoped, reversible, and supported by proof. The reviewer stays focused on the current outcome, and a human still controls the merge.

Fast shipping creates drift

affiliate-mcp has several distribution paths. The npm package carries the server, the plugin path carries skills and registration, and the MCP bundle serves Claude Desktop. A source change can work correctly while one release artefact or version reference stays stale.

We define an authority order for documentation: accepted decisions, policies and runbooks, shipped code and structured manifests, then plans and research. Generated documents are changed through their source. This keeps agents from following different documents that both look current.

Branch scheduling matters too. One pull request actively awaits my review while other work stays queued or in draft. Decisions and foundations land before dependent implementations. Release checks inspect the artefact users will receive.

What I have learned

Coding agents create work faster than a maintainer can safely absorb it. The practical gains come from bounded outcomes, persistent coordination state, repository procedures, independent review, clear decision rights, and disciplined releases.

The procedures still need judgement and regular pruning. Agents can do a remarkable amount inside clear boundaries. Building and steering those boundaries remains part of the engineering work.