Pimp My IDE / garage dispatch
Back to garage
October 2, 2026 | Agents / repeatable work

Put the loop in code.

A long prompt can describe a process. Code can own its fixed stages, typed handoffs, review stops, and run receipts.

Use an agent where judgment is needed. Use code where the same rule must survive the next run.

Parallel agents are not a process.

GitHub's new dynamic workflows put a program around agent work. The workflow author defines stages, conditions, and handoffs. A run can use commands, tools, service calls, one agent, or several agents. Stages can run in sequence or in parallel.[1]

That differs from asking a model to invent the plan on every run. GitHub's documentation separates dynamic workflows from Autopilot and Fleet. Autopilot keeps choosing next steps. Fleet encourages parallel delegation. A dynamic workflow follows a process written in code, although agent answers and input-dependent branches can still vary.[2]

The upgrade is not more agents. The upgrade is moving repeated control decisions out of prose.

Put judgment at a named joint.

Do not encode a twenty-step ceremony because code is available. Start with the parts that must repeat: collect the same inputs, validate the same shape, apply the same limit, pause at the same risk boundary, and save the same receipt.

Use an agent for the work that needs interpretation. Ask it to classify a failure, compare competing explanations, or review a change. Give its output a schema before the next deterministic stage consumes it. GitHub documents structured stage results and can ask an agent to correct a malformed result.[2]

A stop is not an undo button.

Review stops need a visible consequence. The VS Code chat documentation separates queue, steer, and stop. It also states that stopping does not reverse completed file edits, terminal commands, or external actions. A workflow checkpoint should therefore sit before the costly or irreversible step, not after it.[3]

The interface needs to announce state changes too. MDN recommends a polite live region for updates that matter but should not interrupt the current task. The clutch below uses one for its verdict and copy result.[4]

  1. Name the input and fixed stages.
  2. Define a structured result at every agent handoff.
  3. Place a review stop before an external side effect.
  4. Set a run limit for time, retries, and model use.
  5. Save the revision, inputs, stage results, decision, and final effect.

The clutch drafts that operating card. It does not create or run a GitHub Copilot extension.

Interactive makeover / dynamic workflow clutch

Engage repeatable control.

This replaces a vague "automate it" checkbox. Four native switches connect the process rail and write a copyable workflow card with missing implementation evidence left open.

Select card sections

These controls define a template. They do not install an extension, run a workflow, spend credits, or approve an effect.

Workflow interlocks
Process transmission

Clutch disengaged

0 of 4

Workflow card incomplete.

Select the fixed stages first. Later sections stay disconnected until every earlier section is selected.

A completed display means the card has four sections. It does not mean a workflow exists, the extension is trusted, the run passed, or a side effect was approved.

Workflow operating card

The final state is template structure ready. Every bracketed value still needs a real implementation or run result.

Sources read

Source log and evidence boundary
  1. GitHub Changelog, "Dynamic workflows in Copilot CLI and the Copilot app", published October 1 and read October 2, 2026. This supplies the release scope, preview status, stage model, parallel execution, checkpoints, and example uses.
  2. GitHub Docs, "Dynamic workflows", read October 2, 2026. This supplies the distinction among dynamic workflows, Autopilot, and Fleet. It also documents code-defined steps, structured results, direct CLI runs, and input-dependent paths.
  3. Visual Studio Code documentation, "Send messages while a request is running", read October 2, 2026. This supplies the queue, steer, and stop distinction and the warning that stopping does not undo completed or external actions.
  4. MDN, "ARIA: aria-live attribute", read October 2, 2026. This supplies the live-region behavior used for dynamic status feedback in the clutch.

Evidence boundary. We read the release note and documentation. We did not enable Copilot's experimental setting, create an extension, submit a workflow, spend credits, or verify a run. The interactive clutch produces a planning card only.