Lost context
Project memory is versioned data, not whatever still fits in a chat window.
The engine behind the paid machine
Agents move fast.
The workflow should hold.
Bounded Agent Harness is the deterministic control plane underneath Supercharge: versioned memory, contract-checked handoffs, narrow authority, bounded cost, and recovery you can inspect. When a bundle ships with the full harness, this is what runs it.
workflow / review-releaselocal proof
sha256:7590…7dec0
plan.validatecontracts · authority · cost
specialist.reviewmcp · bounded context
artifact.acceptevidence · checkpoint
continuous run activecost 7 / 8 units
Verified in the release candidate. Reproducible evidence, clearly scoped.
191passing tests
0runtime dependencies
READYproof-release gate
Current automated result on Linux: 191 passed, zero failed, zero skipped.
Loading test report…
The control loop
Less agent theater.
More operational truth.
- 01
Validate
Contracts, authority, capability, budget
- 02
Compile
One deterministic workflow identity
- 03
Execute
MCP adapters with token-capacity gates
- 04
Continue
Continuous run; heartbeat monitors and recovers
Built for the failure around the model
Models are capable.
Operations are fragile.
Bounded Agent Harness does not replace your models or tools. It gives them a smaller, inspectable surface to work inside.
Loose authority
Every specialist is filtered by capability and allowed actions before assignment.
Brittle handoffs
Inputs and outputs cross named, fingerprinted contracts or fail closed.
Stopped after status
A progress report is not terminal. Safe work continues in the active run; heartbeat only monitors and recovers.
Verified capabilities
A tested local control plane, with the next integration step visible.
Version 0.4.0 compiles, runs, resumes, inspects, and rejects unsafe local workflows. It includes contract-bound memory, evidence, cost enforcement, approval gates, specialist routing, portable takeover checkpoints, an in-process MCP compatibility proof, a token-capacity gate, an additive token-efficiency layer for lazy capability loading, digest-bound artifact references, deterministic context pruning, and measured cache savings.
Direct proof on this Linux host: project-scope validation passes 10/10 tests; source materialization passes 11/11, including symlink-path rejection; controlled execution passes 19/19. The controlled API requires declared context and versioned acceptance before advancement when the host invokes it.
Proof scope
- Local literal adapters; no model or network calls
- 21 ledger events; one rejection and explicit resume
- Declared context and independent artifact checks
- 11 cost units, including the failed attempt
Known limitations
What this doesn't do.
- The /try demo replays a recorded run — no live model calls, no spending, no surprises. That's the point: the control plane is inspectable before you trust it.
- The engine proof runs locally with literal adapters. No model calls, no network calls.
- The Supercharge bundle is prompt discipline you paste into your own AI — not a hosted runtime, not a new app, not a subscription.
- Hash chaining detects inconsistent edits; it does not authenticate a ledger against complete rewriting.
- The harness governs work; it does not replace your models or tools.
Used in practice
From harness to a real Site build.
The USB Website Platform used Bounded Agent Harness while building and verifying deterministic Site artifacts for DFW Metro and the disposable DFWMow review lab. The harness carried shared task packets, scope reservations, handoffs, checkpoints, focused proof, and rollback awareness; production Site generation itself remained deterministic and model-free.
Resume, don't fail
Missing or stale approval becomes a durable checkpoint with the exact gate and workflow identity required to continue later.
Spend proof where it matters
Change-impact cards choose focused verification so documentation and local edits do not trigger needless full-suite work.
Keep safe work moving
A blocked gated action pauses that action; unrelated local work can continue until a real boundary or proven completion.
Real integration
The harness was mounted during a deterministic Site build, sidecar proof, deployment preflight, and same-origin pilot; focused tests passed and live readback was recorded.