Harness Decision Worksheet

Use one row for each workshop exercise. The goal is not to collect every possible feature. The goal is to justify each component with evidence.

Run identity

Exercise and lane:
Live run or saved-evidence audit:
Workspace:
Date and software versions:
Resolved model and reasoning:
Enabled tools, memory, and skills:
Limits enforced by the harness:
Limits requested only in the prompt:
Provider usage source, or unknown:
Independent checker command and exit status:

What changed?

The harness choice I tried:

What I expected:

What I noticed in the trace:

What the optional check told me:

What I would change for another run:

Decision	Our choice	Evidence we expect	Cost or new failure mode	Escalation trigger
Working trace
Tool surface
Scaffolding and context
Model routing
Long-running control
Multi-agent design

Opening Trace

Our selected task:

We predicted the largest difference would be:

I trust the result because:

The harness, rather than the model alone, supplied:

The first efficiency measurement I would add is:

The trace explains:

The provider ledger measures:

If these disagree, I will:

After an interruption, the first thing I would reverify is:

Tool-Surface Rule

Use the terminal or API when:

Add an MCP connection when:

Load a tool by default when:

Make it discoverable on demand when:

Prefer a small API or CLI interface when:

Block or require approval when:

The evidence that would change our choice is:

Context Budget

A line earns a place in AGENTS.md when:

A procedure becomes a skill when:

A fact earns persistent memory when:

Keep state in a task checkpoint when:

Retrieve information again when:

Enforce a rule outside the model when:

Delete saved context when:

We will review and prune these artifacts when:

Long-Running Contract

Select the state that must survive an interruption:

• [ ] Objective
• [ ] Explicit completion criteria
• [ ] Current plan and next action
• [ ] Changed files and repository state
• [ ] Latest verifier evidence
• [ ] Failed approaches
• [ ] Environment identity
• [ ] Remaining token, cost, or iteration budget
• [ ] Allowed actions and workspace boundary

Goal:

Completion requires:

Progress requires new evidence such as:

Continue when:

Change strategy when:

Reverify after:

Escalate when:

Stop when:

Maximum turns, time, tokens, or cost:

Model Routing Matrix

Task	Initial model or tier	Reasoning level	Verifier	Escalation signal
Simple edit
Off-by-one bug
Async bug
Authentication review

Multi-Agent Boundary

Another agent earns its cost when:

The work can be separated because:

Keep one agent when:

Use an independent agent as validator when:

Use code or another deterministic gate when:

Each child owns:

The child returns:

Agents may share:

Agents must not share:

The synthesis owner is:

The independent verifier is:

The total compute budget is:

Stop parallel work when:

Final Recovery Plan

List the next three actions in order. Each action should cite state, evidence, risk, or budget.