Harness Atlas

Two plates from one survey of an agent harness: how it routes a task across roles, and how it keeps whatever model you plug in inside safe bounds.

safe structure active thread rejected step not yet traced
I

Role Topology

dispatcher · 4 ARCH 1 idle WORKER 2 idle EXECUTORS 3 idle REVIEWER 5 idle ×0 IMPL-FIXER 5 idle ×0 6a 6b APPROVED
A central dispatcher moves one task through five agent roles. Review and fixing are loops: they repeat until the change passes.
scenario
speed
step 0 / 0

  1. 1ARCHdraft the plan—
  2. 2WORKERimplement the change—
  3. 3EXECUTORSbuild in parallel—
  4. 4Dispatchermerge the results—
  5. 5REVIEWERcheck against the plan—
  6. 6Outcomeapprove, or loop back—
II

Guarded Corridor

LLM containment limit Wall Shield protection limit keeps the agent in keeps the workspace safe goal
Every step is one model call. Two limits hold the run: a wall that keeps the agent in, a shield that keeps the workspace safe. A step that would cross either one is rejected before it runs.
scenario
speed
step 0 / 0

attempt log