Skip to content
neℓson

ἓν οἶδα ὅτι οὐδὲν οἶδα

Ideas

A delivery loop for agentic engineering

A spec-first agent delivery loop keeps speed useful by preserving review points, validation signals, and production ownership.

This is a spec-first delivery loop for using agents without giving up engineering judgment.

The useful part is not that an agent can write code quickly. The useful part is that the work moves through explicit checkpoints where risk, validation, and ownership stay visible.

The operating loop

flowchart TD
  A["📝 Spec Draft"] --> B["⚖️ Human Review 1 Risk Scope"]
  B -->|Reject| A
  B -->|Approve| C["🌿 Spec Branch"]

  C --> D["⚙️ CI Runner"]
  D --> E["🤖 Agent Execution"]

  E --> F{"🧪 Validation Checks"}

  F -->|Fail| G["🔁 Auto Feedback Patch"]
  G --> D

  F -->|Pass| H["📦 Pull Request"]

  H --> I["👨‍💻 Human Checkpoint 2"]
  I -->|Changes| G
  I -->|Approve| J["🚀 Merge Main"]

  J --> K["📊 Post Merge CI"]

  K --> L{"🚦 Production Ready"}

  L -->|Fail| M["🧯 Rollback Patch"]
  M --> C

  L -->|Pass| N["🎯 Complete"]

  N -->|Next| A

  class A,C spec
  class B,I human
  class D ci
  class E agent
  class F,L validation
  class H,J pr
  class K post
  class G,M loop
  class N done

  classDef spec fill:#1e3a8a,stroke:#3b82f6,color:#ffffff
  classDef human fill:#334155,stroke:#94a3b8,color:#ffffff
  classDef ci fill:#581c87,stroke:#a855f7,color:#ffffff
  classDef agent fill:#6b21a8,stroke:#c084fc,color:#ffffff
  classDef validation fill:#92400e,stroke:#f59e0b,color:#ffffff
  classDef pr fill:#065f46,stroke:#10b981,color:#ffffff
  classDef post fill:#7f1d1d,stroke:#f87171,color:#ffffff
  classDef loop fill:#111827,stroke:#9ca3af,color:#ffffff
  classDef done fill:#14532d,stroke:#22c55e,color:#ffffff

Why the first review matters

The first human checkpoint is not a code review. It is a risk-scope review.

Before the agent starts, the spec should make the blast radius legible: what can change, what must not change, and which checks prove the work is acceptable. Rejection at this stage is cheap because no implementation has been produced yet.

Where agents fit

Agent execution sits behind CI instead of replacing it.

The agent can patch, rerun, and respond to validation feedback, but the loop is bounded by tests, type checks, build output, and reviewable diffs. That keeps speed tied to evidence rather than confidence.

Why there is a second checkpoint

The pull request checkpoint is where a human judges whether the result is maintainable, not only whether it passes.

If the change needs adjustment, it goes back through the feedback patch loop. If it is approved, it merges into main and faces post-merge validation as a separate production-readiness gate.

Completion is temporary

The loop ends by returning to the next spec draft.

That matters because agentic engineering is not a one-shot generation event. It is a way to keep delivery moving while preserving the review, rollback, and production checks that make software dependable.