01
Start here
Before the first flight: scope, budget, kill switch
What to set before an AI agent does real work: a deliverable you can judge, permissions that match the scope, a time and spend limit, and how to stop it.
The Field Manual
Everything we know about holding the reins.
Practical references for the decisions every rider makes: what to delegate, where to draw the boundary, and how to know the work is done.
-
-
02
Reins: permissions that are tight enough to be loosened
AI agent permissions, enforced rather than instructed: least privilege, approval before the side effect, retrieved data treated as data, not instructions.
-
03
The bond: how to give an agent more room, one flight at a time
Four stages of AI agent autonomy and how to earn each: trust the setup, not a personality, change one boundary at a time, and keep a promotion record.
-
04
Burns: what goes wrong, what it costs, what to do in the first ten minutes
What an AI agent incident costs and the first ten minutes: contain the run, establish the facts, choose the smallest recovery. Riders call it a burn.
-
05
Logging a flight: the template and why every field matters
What to record for each AI agent run: model, permissions, elapsed and human time kept separate, a supported cost, the outcome, one change for next time.
-
06
Riding a wing: running several agents without losing the reins
Multi-agent work without losing control: divide by deliverable, give each handoff a contract, keep permissions narrow, budget the whole fleet.
-
07
Evals: how a rider keeps a dragon accountable
How to evaluate an AI agent's work with repeatable checks: golden sets, rule-based assertions, LLM-as-judge, human spot-checks, and a tracked pass rate.
FIELD MANUAL / GUIDES
Reference guides
Standalone references that sit beside the chapters: how to stop a run, how to cap what it spends, and how the four mounts compare.
-
Kill switch for an AI agent
A kill switch for an AI agent is a stop condition plus a mechanism outside the model that enforces it. Five mechanisms work: a process or session kill, a hook that blocks the tool call, an approval gate before side effects, a turn or budget cap, and revoking the credential. Fire each on a harmless run before trusting it.
-
Cap what a flight can spend
A spend cap stops an AI agent at a limit; an alert only tells you afterwards. Caps count tokens, dollars or turns, per run or per day. Claude Code has a per-run dollar cap in print mode; Codex and Gemini CLI do not, and Cursor documents none, so their enforceable limit sits in billing, turns or a hook you write.
-
Four mounts, four sets of reins
AI agent permissions, run caps and kill switches differ across Claude Code, Cursor, Codex CLI and Gemini CLI. Claude Code has allow, ask and deny rules and a per-run dollar cap in print mode. Codex sandboxes by default and fails closed. Gemini CLI ships a turn cap and a policy engine. Cursor uses a classifier, hooks and team spend limits.
THE NEXT HORIZON
Put the craft into practice.
One task. One outcome. One honest record.