Durable by design
State lives in a journal, not only in memory. Work can suspend, survive a crash, and resume without repeating completed steps.
Project case study · Codename Truss
Agent work you can verify
without reading the transcript.
Models are increasingly capable of doing the work. The harder problem is deciding whether the result deserves your trust. I built a provider-independent agent harness that accepts a contract, executes it durably, verifies the outcome outside the executor, and returns signed evidence.
- Unit of work
- A contract, not a prompt
- Acceptance
- Independent checks and review
- Output
- A signed, portable verdict
The execution path
Contract in. Proof out.
The executor never grades its own homework. Completion and acceptance are separate jobs with separate evidence.
01 / CONTRACT
Define done first
Intent, machine-checkable completion criteria, reviewer criteria, budget, and constraints are fixed before execution.
02 / EXECUTE
Use the right worker
Run an in-process model, a local model, or delegate to an existing coding-agent subscription behind one contract.
03 / VERIFY
Check independently
The harness runs deterministic checks and adversarial review in contexts separate from the executor.
04 / VERDICT
Keep the receipt
The run, evidence, and verdict live in a hash-chained, signed ledger that can be verified later.
Why build it
The bottleneck moved.
The industry keeps optimizing how agents generate. This project starts from a different premise: as generation gets better, governance becomes the limiting system.
Trust should be an artifact of the run, not a feeling after reading it.
Provider-independent core
The runtime does not belong to one model vendor. Provider and executor adapters can change without changing the trust path.
Policy at the boundary
Budgets, deadlines, tool permissions, and filesystem scope are enforced where actions happen—not left as suggestions in a prompt.
Private evidence
Run content stays local. Tamper evidence can be anchored externally using hashes without publishing the underlying work.
What this demonstrates
A working reference architecture for governed agents.
This is an engineering project, not a renamed chat interface. Its components map directly to the hard parts of deploying agents inside real organizations.
Orchestration
One contract can govern different executors while keeping budgets, constraints, and acceptance criteria consistent.
Verification
Deterministic checks and independent review produce evidence tied to the original definition of done.
Provenance
Signed, hash-chained events make the history inspectable and make later tampering detectable.
Local-first operation
Local and open models can stay inside the company boundary, with frontier endpoints used only when the task requires them.
Evaluation
Paired runs and benchmark adapters measure whether changes improve pass rate, cost, and user outcomes.
Integration seams
Providers, tools, context sources, and event consumers attach through explicit boundaries instead of leaking into the core loop.
Bring the pattern to your workflow
You may not need this harness. You probably need these guarantees.
I can apply the same contract, policy, verification, and audit architecture to the AI workflow your team is actually trying to ship.
Start a project