Five ways an agent can fake a green build
Agents rarely cheat on purpose, but under pressure to make tests pass they find shortcuts. Here is what to look for — and what PaneScale flags automatically.
What the next agent needs to know
When one agent gets stuck and another takes over, the handoff decides whether the second attempt is better or just different. Anatomy of a continuation brief.
Agents get stuck. They run out of attempts on a hard bug, they hit a usage limit halfway through a refactor, they circle around a test they don’t understand. Often a different agent gets further — different training, different habits, a fresh look.
But a fresh look is only useful if it doesn’t start from zero. The second agent needs to know what happened. In PaneScale that information travels as a continuation brief.
What goes in
The original task
Not a summary of it — the task itself, with its finish line. Summaries drift; each retelling loses a constraint.
A plain statement of the situation
“Another agent already worked on this task in this directory and got stuck. Its changes are still here.” This one sentence changes the second agent’s behavior completely. Without it, agents tend to treat existing changes as the starting codebase and build on top of mistakes. With it, they inspect first.
An instruction to keep what is good
The previous agent rarely got everything wrong. Asking the next one to review the existing changes, keep what is good and fix what isn’t is cheaper than starting over — and usually produces a smaller diff.
The last failing check
The exact output of the check that failed last. This is the most valuable part: it tells the next agent where reality disagreed with the previous attempt.
What stays out
- The previous agent’s reasoning. It is long, and it is the reasoning that got stuck. The diff and the failing check carry the useful part.
- Opinions about the previous agent. “The last agent made a mess” primes the next one to throw everything away.
Where handoffs happen in PaneScale
- The Overnight Autopilot runs one lead agent with retries and hands each task to a backup agent for the second half of its attempts.
- From any pane, choose Hand off to another agent… and PaneScale writes a brief from the pane’s session for the agent you pick — see Handoff.
- Between teammates in Agent mode, a teammate can pass a subtask to another one — see Agent mode.
The autopilot and pane handoffs keep the work in the same place. The next agent opens the same files the last one touched — no copying, no pasting, no lost changes.
Choosing the backup
The best backup is the agent most different from the lead. If your lead is good at broad changes, pick a backup known for small, careful steps — or the other way round. A handoff between two agents with the same strengths tends to get stuck in the same place.