Guides/Agent boundaries & HITL

Agent boundaries & human-in-the-loop

Agents fail loudly when tools do side effects without governance — PMs ship confidence by tiering tools by blast radius, pairing automation with approvals, and logging decisions reviewers can audit.

Action lanes

Model capability matters less than which actions run unattended. Align lanes with security + legal — not only ML metrics.

Map tools into tiers before choosing models — governance surfaces often ship inside policy engines, not the LLM weights.

Risk tiers → checkpoints

Adapt tiers to your regulatory posture — use this table to spark RACI-style conversations before coding tool routers.

Risk tier	Example surfaces	Human-in-the-loop pattern
Low	Summaries, suggestions, read APIs	Automated metrics + spot checks
Medium	Customer-visible drafts, internal writes	Human review queue or dual-control on odd hours
High	Money movement, HIPAA/FINRA paths, deletes	Hard approval, break-glass, immutable logs

Operating principles

Default-deny tool catalogs — promote capabilities through review instead of blocking after incidents.
Pair agent launches with rollback + replay affordances for support teams.
Treat prompts as insufficient guardrails — enforce policy in orchestration layers humans can inspect.