Skip to content

Playbook

Agentic workflow for coding

Autonomy without a loop is chaos. This is the process shape that keeps multi-step agents shipping software you can defend.

Default loop

Specify → plan → implement → verify → review → merge. Every arrow is a place an agent can fail. Verification is the only arrow that must never be owned by the author-agent alone.

The loop in practice

  1. 01

    Specify

    Acceptance tests, out-of-scope edges, and risk files named up front.

  2. 02

    Plan

    Read-only scout produces a bounded plan with file list.

  3. 03

    Implement

    Builder edits in a worktree or branch. Keep diffs reviewable.

  4. 04

    Verify

    Real build and test commands. Exit non-zero means continue with evidence.

  5. 05

    Review

    Different model or human on auth, data, and deploy surfaces.

  6. 06

    Merge

    Only after gates pass. Checkpoint so rewind is cheap.

Patterns worth knowing

Fan-out

Parallel independent tasks. Fails when files collide.

Pipeline

Same stages across many similar items.

Reviewer-gate

Different lab reviews. Gate needs teeth.

Loop-until-dry

Exhaustive hunts with an objective empty state.

Full pattern writeupAI agent orchestration patterns. Team rolesmulti-agent field guide.

Tools that run the loop

FAQ

What is an agentic workflow?

An agentic workflow is a defined multi-step loop where AI agents act with tools toward a goal, and humans set scope plus verification. Typical stages, plan, implement, test, review, merge.

What are common agentic workflow patterns?

Fan-out for independent tasks, pipeline for staged processing, reviewer-gate with a different model, and loop-until-dry for exhaustive cleanups. All need external verification.

How is this different from vibe coding?

Vibe coding is the intent style. An agentic workflow is the process that keeps autonomous steps coherent and shippable.

Do I need multiple models?

Not always. A different model as reviewer is high leverage. Multi-model crews help when labs share blind spots or price/performance differs by role.

Where does a harness fit?

A harness runs the workflow across CLIs, roles, isolation, shared memory, and a verify gate that is not the author-agent’s self-report.

Workflow as product

VibeSwarm is this loop with real CLIs

Coordinator, builders, reviewer, AutoVibe gate — your keys, your tests.

The app behind this research

TheVibeFather is the multi-CLI AI coding harness

Run Claude Code, Codex CLI, OpenCode, Grok Build and other agents as one crew — shared project memory, roles, and a verify gate so your tests decide what ships. Bring your own keys.