Operating manual
Tag: AI Agents
A focused reading path for AI Agents: related field notes, evidence trails, and operating questions from the archive.
- self-correction-needs-executable-verification.md
AI Self-Correction Needs Executable Verification
A practical verification promotion gate for deciding when an AI agent correction has enough evidence to replace prior state or trigger action.
open artifact → - validate-ai-agent-trajectories.md
How to Validate AI Agent Trajectories
Component tests prove local properties. Deployable AI agents need evidence for the path they take across state, authority, tools, memory, and handoffs.
open artifact → - agent-simulations-need-a-reproducibility-contract.md
Agent Simulations Need a Reproducibility Contract
A seed is not enough. Reproducible agent simulations need a versioned contract for execution, dependencies, validation, and rerun evidence.
open artifact → - simulation-boundaries-need-infrastructure.md
Simulation Boundaries Need Infrastructure
A practical four-control model for testing whether an AI agent sandbox truly restricts network access, credentials, side effects, and recovery.
open artifact → - more-compute-wont-fix-local-computer-use-agents.md
Four Compute Bets for Computer-Use Agents
For computer-use agents, context, steps, decomposition, and parallel plans are different bets. Here is how to test which one earns its cost.
open artifact → - confidence-should-not-allocate-your-audit-budget.md
Confidence Should Not Allocate Your Audit Budget
Agent confidence is one audit signal, not the allocator. Use a five-factor scorecard and random reserve when human review capacity is scarce.
open artifact →