Two agents in two threads share files but not context. Both decide on stale state. Fix: one fresh agent per task, with isolated context.
Workflow
Structured approaches to AI-assisted dev — spec-driven development, planning skills, brainstorm-before-code, OpenSpec and OPSX artifacts.
The reliability problem is not the model. It is the missing disciplines around it. Brainstorm, plan, test, verify, review: each gate before the next step.
A ten-task plan drifts by task four. The model is not forgetful. The plan is too coarse. Fix: smaller tasks with sharper edges.
Ten or more active worktrees with no naming and cleanup rules becomes a graveyard of old branches and lost work. Three rules fix it.
An agent that runs forty minutes cannot survive a laptop sleep. The fix is a session that outlives the laptop, not faster agents.
Spawning agents is easy. Reviewing their autonomous decisions without a logging contract is the bottleneck. Here is the format that fixed it for me.
The default workflow will not fit your team forever. You can customize it at three levels. Most teams should stop at level one: the project config.
Specs go stale because updating them hurts. A delta fixes this: describe changes in three sections, ADDED, MODIFIED, REMOVED, with no full rewrite.
You inherited a codebase with no specs, so the agent guesses. Fix: spec going forward, not backward. One change at a time, thirty-minute setup.
Phase-locked workflow blocks you from going back when you learn mid-task. The fix: actions you run anytime, in any order, without restarting the change.
A change folder has four files: proposal, specs, design, tasks. Teams guess which to update, and artifacts drift. The fix: know what each is for.
You give the agent a one-sentence task. It codes right away. The code looks fine but is not what you wanted. Fix: brainstorm before code.

