01 · GOAL
Goal boundary
Define the target outcome narrowly enough that completion can be recognized.
LESSON 02 · BOUNDARIES
Define the Agent Boundary
An agent should know what it is trying to accomplish, what it may know, which tools it may use, and exactly where it must stop.
Bound the system before you expand it.
Agency becomes risky when the system can pursue a broad goal with unclear limits. A production agent needs explicit boundaries across seven dimensions: goal, knowledge, tools, authority, time, cost, and failure handling.
D→V RULE
A boundary is useful only if the runtime can enforce or detect it. “Be careful” is not a boundary. “Do not call change_booking without approval_id” is.
02 · KNOWLEDGE
Knowledge boundary
Specify approved sources, freshness requirements, and what the agent must treat as unknown.
03 · TOOLS
Tool boundary
List permitted capabilities and forbid implicit access to adjacent systems.
04 · AUTHORITY
Authority boundary
Separate read, recommend, modify, commit, spend, communicate, and delete.
05 · TIME
Time boundary
Limit run duration, stale state, retry windows, and long-running loops.
06 · COST
Cost boundary
Define spend, token, tool-call, and external-service limits.
07 · FAILURE
Failure boundary
Describe when the system stops, abstains, compensates, or escalates.
OUTSIDE
Forbidden
- Unapproved data sources
- External commitments
- Deletion or irreversible action
AGENT OPERATING ENVELOPE
Investigate shipment exception
- Read approved operational evidence
- Form hypotheses
- Select permitted investigation tools
- Recommend recovery options
HUMAN GATE
Requires approval
- Change a booking
- Issue customer credit
- Send an external commitment
INTERACTIVE LAB · BOUNDARY CHALLENGE
Should the agent proceed?
Evaluate actions against the operating envelope. The correct answer depends on the boundary, not on whether the action sounds useful.
CASE 01
Read terminal status
The agent wants to query an approved terminal-status API to verify congestion.
CASE 02
Change the booking
The agent has found a faster route and wants to confirm the new booking with the carrier.
CASE 03
Search an employee mailbox
Shipment evidence is incomplete, so the agent proposes searching unrelated employee email accounts for more context.
CASE 04
Evidence conflict
Carrier ETA and terminal-status records disagree materially, and the customer decision window closes in 15 minutes.
LESSON ARTIFACT
Build the Agent Boundary.
Write operational limits that can later be enforced by permissions, policies, runtime checks, and evaluation cases.
Sample · freight exception agent
- GOAL
- Investigate late shipments and recommend recovery options before the decision window closes. Do not make customer or carrier commitments.
- KNOWLEDGE
- Approved shipment events, weather, terminal status, SLA, and contracted recovery options. Treat missing or conflicting evidence as unresolved.
- TOOLS
- Read-only shipment, weather, terminal, SLA, and route-cost tools. No direct booking mutation.
- AUTHORITY
- Read, analyze, and recommend. External communication, financial commitments, and booking changes require approval.
- TIME
- Maximum 10-minute investigation; no evidence older than the operational freshness threshold without warning.
- COST
- Cap tool calls and model spend per case; escalate rather than loop indefinitely.
- FAILURE
- Stop on material evidence conflict, missing mandatory data, tool authentication failure, or any proposed action outside the approved authority.