Harness Engineering
Agent systems
I started by building a governance framework around OpenClaw. My work now spans Hermes, Pi, DeepSeek Harness and model routers: keeping project intent and verification steady while changing how agents run and which models they use.
Ongoing harness engineering
OpenClaw made the boundaries explicit.
My OpenClawGovernor work separated the operator’s intent from runtime execution. A Governor outside the fleet managed configuration and correction; an orchestrator delegated to domain directors and workers. Specs and guard plugins made that separation concrete. The enduring lesson is to keep authority, context and responsibility clear. I am carrying that lesson forward without requiring every new project to reproduce the same hierarchy.
A useful boundary should make work easier to reason about. A permanent extra layer needs to earn its place.
- Input
- An objective, permitted scope and written specification
- Output
- A bounded piece of work with a clear owner
- Mechanism
- Intent → scope → delegated work → review
- Technology
- OpenClawGovernor · specifications · role and announce guards
Keep the project contract outside the harness.
An agent harness supplies the execution loop, tools and working context around a model. I want the project’s brief, task boundaries and acceptance checks to remain understandable when that harness changes. ProjectHarness explores how to keep those responsibilities separate.
Portability comes from explicit interfaces. Session formats, permissions and tool behaviour still need an adapter and testing.
- Input
- A task, its dependencies and an executable acceptance check
- Output
- A checked result with status and failure evidence
- Mechanism
- Project record → ready task → worker → verification
- Technology
- ProjectHarness · Ringer/Ringside · task state · executed checks
Change the model without redesigning the work.
I am using Hermes more and exploring Pi and DeepSeek Harness alongside the earlier OpenClaw setup. OpenRouter and Claude-compatible routers make it easier to try a different model within an existing workflow. Each combination still has to prove its tool use, context handling and finished result on the actual task.
The model, its provider and the harness are different variables. Record them separately so a model experiment produces useful evidence.
- Input
- The same task and checks, with a selected harness and model
- Output
- Comparable attempts with the execution choices recorded
- Mechanism
- Task → harness → model route → checked result
- Technology
- Hermes · Pi · DeepSeek Harness · OpenRouter · model routing
Use parallel agents where the work can be separated.
The aim is a small coordination layer around focused workers, with a shared project record and explicit handoffs. Independent research, implementation and review can run alongside one another; dependent changes wait for their inputs. My wider projects reinforce this approach: Spec-first keeps intent connected to implementation, NewsApp Windmill delegates execution and recovery to Windmill, with Jev / TypeSafe System One handling typed classification and local Qwen generating prose, and ProjectHarness explores dependency-aware delegation and executed verification.
Acceptance comes from inspecting artifacts and running checks. Keep failures visible and retain human judgement for what those checks cannot decide.
- Input
- Independent tasks with bounded context and explicit dependencies
- Output
- Checked artifacts, recorded failures and a decision about the next step
- Mechanism
- Delegate → execute → check → integrate
- Technology
- Scoped workers · dependency order · acceptance checks · human review
Recorded evidence
Mike Lowe · working practice, 9 September 2026 · repository evidence
- 01 Apr 2026 · Template aligned with fleet rebuild · Committed · 156f332
- 02 Apr 2026 · Announce and role guard plugins · Committed · c565127
- 02 Apr 2026 · Guard deployment added to setup · Committed · e70c880
- 10 Jul 2026 · ProjectHarness fork adds an orchestration layer · Local commit 1f713a4 · builds on Nate B. Jones’s Ringer/Ringside; live Linear smoke testing remains outstanding
- 09 Sep 2026 · The practice broadens beyond OpenClaw · Current working direction · more Hermes, Pi and DeepSeek Harness exploration, with model routing and simpler delegation
Harness Engineering
Related work