Skip to content

✅ Dev Log 034 – Agent Plans Get Real

We spent the last few focused sessions refactoring the dokugent plan command — not just to make it better, but to make it composable for dry runs and simulations.

Until now, the plan wizard felt like a placeholder. It worked, but it didn’t hold weight. It asked you what the agent should do, but didn’t respect the structure of actual planning.

That’s changed.

What’s New

  • Each plan step is now its own file. Stored under .dokugent/data/plans/<agentId>/steps/<stepId>.md, every step is a unit of intent: with goals, roles, tools, and constraints.

  • The wizard is no longer naive. It skips step IDs you’ve already defined. It warns you before overwriting. It uses the active agent as context, and doesn’t make you repeat yourself.

  • plan.md is assembled, not authored. Steps are written individually, but Dokugent automatically compiles them into a full plan. Every update rebuilds it. Nothing gets lost.

  • Token count feedback is built-in. After every run, we estimate token cost. You know what your plan is really costing before you simulate or dry run.

  • It feels more like an operating system now. Not in a bloated way — in a minimal viable control surface kind of way. Like something you could trust to scaffold a real agent.

Why It Matters

Plans are where agents become accountable. If an agent fails, we need to know what it was supposed to do, step by step. That means those steps have to be inspectable. Composable. Auditable.

This refactor makes that real.

What’s Next

  • dokugent plan --ls and --preview are next.
  • Certification will eventually read these steps as structured input.
  • Simulation will let you mix and match combinations without rewriting anything.

This DX polish is groundwork for traceable agent workflows.

More soon.

Categories