The Push Bike Harness
Three approaches to teaching someone to ride a bike.
Training wheels: the bike stays upright. No falls. Also no balance — the learner is balancing the training wheels, not themselves. Removing them later requires relearning from scratch. The protection becomes the obstacle.
Bubble wrap: the bike and the kid.
Push bike: no pedals, just the learner and the balance problem. There are falls. The balance develops anyway. When pedals arrive, they’re trivially easy — the rider already knows how to stay upright.
Most AI agent harnesses are training wheels. Heavy pre-execution gates, blocking checks, extensive sandboxing — all legitimate tools for specific problems. But if the harness prevents every fall, it may also prevent the learning that comes from falling. Some falls are how you find out where the road is rough.
DEAR is my attempt at a push bike design. Four phases: Define (specify what “done” means before starting), Enforce (block known failure modes at execution time), Audit (catch what slipped through), Resolve (fix it — not just the instance, the root cause). The loop is closed: R feeds back into D and E, making the harness permanently better after each failure.
The key design principle is Deferred Enforcement: not every check belongs in the synchronous path. If the cost of checking synchronously exceeds the cost of catching and fixing it later, skip the gate and rely on Audit. When 17 research sessions stranded or disappeared — 10 on unmerged branches, 7 cleaned by the reaper — the fix wasn’t more gates. It was a better definition of “complete” (five machine-checkable predicates), a pre-exit hook that blocked termination if they failed, and a periodic audit that caught stranded branches before cleanup ran. Sessions after that: zero lost. Same road, smoother surface.
I should be honest about where the harness is. The Enforce layer is solid — hooks, circuit breakers, branch tracking. The Audit layer is partial — the Overseer runs, the event log exists, cross-harness verification works. The Resolve loop, the part that automatically improves the harness from its own failures, is the most aspirational piece. The plumbing is sketched. The automation isn’t there yet.
And the harness itself has training wheels that probably need removing. Five-role orchestration designed for a scale not yet reached. Cross-harness verification on every task, including ones where the cost likely doesn’t justify it. The Deferred Enforcement principle is in the protocol. The harness doesn’t always follow its own advice.
The dear-agent repository is one person’s attempt to operationalize this philosophy. Offered as a starting point, not a prescription.
The ideal harness doesn’t prevent falls. It catches them, root-causes them, and teaches the agents how to prevent or handle them the next time around. After enough cycles, nobody thinks about the guardrails — not because they’re gone, but because the surface is clean.
- Push bike.
- Gotta go fast!
- …?
- Profit
Related reading:
- Don’t Hate the Agent, Hate the Process — on the process failures DEAR is designed to catch
- Most Rules Exist for a Reason — on why the harness itself needs to understand its own rules
- Oops I Did It Again, I Forgot –dry-run — on safety systems that let you go faster