liveBook · Manning

Constraint

As explained in Architecting Autonomous Agentic Systems

A rule or limit that restricts the scope, data, actions, retries, evidence sources, or communications permitted in an interaction.

In an agentic interaction, a constraint turns an open-ended request into a bounded action surface. It limits what the agent may do, which context and evidence it may use, and when it must clarify, reject, escalate, or hand control to a person. section 6.2.2 section 6.2.1

A constraint works with intent, context, authority, expected outcome, and evidence obligation. Together, these elements define why an interaction exists, how it is interpreted, what action is permitted, what completion means, and what record must remain. section 6.2.2

Apply constraints at explicit enterprise decision boundaries:

  • Restrict permitted intents rather than merely naming an API operation; “assess whether lot 2418 requires containment” carries more governance than “query production history.” section 6.2.2
  • Limit actions through delegated authority, required context, runtime policy checks, and an approved operating envelope. section 6.2.2 section 6.2.1
  • Define denial, clarification, escalation, and human-handoff paths when the request exceeds the agent’s authority or lacks sufficient context. section 6.2.1
  • Preserve the contract instance and evidence obligation across APIs, tools, events, workflows, and agent-to-agent exchanges. section 6.2.1

Quality investigation

Elena asks a supervisor agent to assess whether a production lot requires containment. The constraint boundary identifies the affected lot, required production and supplier evidence, permitted recommendations, delegated authority, and the point where Elena, supplier quality, or regulatory affairs must take control. section 6.2.2 section 6.2.1

Figure 6.2 Anatomy of a conversational contract. A conversational contract carries an agentic interaction from intent toward outcome. Context shapes how the intent is interpreted, constraints and authority define the boundary within which the agent may act, and the evidence obligation defines what must remain after the agent has acted.

Use the contract anatomy to trace how constraints and authority bound movement from intent toward outcome while evidence remains after action. section 6.2.2

Table 6.2 Conversation-as-Interface pattern card
Attribute
Definition & Requirements
Mandate
Treat the agentic interaction as a governed contract surface through which intent, context, constraints, authority, outcome, and evidence obligations are made explicit before the agent invokes tools, exchanges work with other agents, publishes events, or hands control to a human.
Scope
Applies where an autonomous or semi-autonomous agent interacts with a person, enterprise system, tool, resource, workflow, event stream, or another agent under delegated authority.
Operating envelope
Valid when the interaction can be bounded by an explicit conversational contract that defines permitted intents, required context, runtime constraints, authority checks, expected outcomes, evidence obligations, and escalation or denial paths.
Boundary of authority
The agent may act only within the intents, constraints, and authority granted by the conversational contract, and only through the interaction types the contract permits. Actions outside the contract must be rejected, mediated through a governed tool wrapper, routed through an approved agent-collaboration path, or escalated to a human or a supervisory agent.
Failure mode
If intent is ambiguous, context is incomplete, authority is insufficient, the requested outcome exceeds the contract, or the interaction type is not permitted, the interaction must fail closed through clarification, rejection, escalation, or handoff rather than continue through implicit interpretation.
Evidence & auditability
The interaction must leave a record of the expressed intent, relevant context, constraints applied, authority basis, interaction mechanism used, outcome requested, actions taken, and any escalation, denial, or handoff path triggered.

Use the pattern card to inspect the required operating envelope, authority boundary, evidence obligations, and escalation paths for a governed interaction. section 6.2.1

How Architecting Autonomous Agentic Systems covers it

How deep it goes
Defines runtime boundaries and decision paths but deliberately leaves internal agent reasoning unspecified.
Examples
Uses a production-lot containment assessment requiring defined records, recommendations, delegated authority, and escalation.
What it assumes
Assumes familiarity with APIs and event contracts while making intent, authority, constraints, and evidence explicit.
What you can do after
Operate/configure governed interactions that constrain actions, preserve evidence, and route exceptions to authorized humans.
How it connects
Connects conversational contracts with intent, context, authority, evidence obligations, APIs, workflows, MCP, and A2A exchanges.

Also explained in

Compare how 29 books explain constraint →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Architecting Autonomous Agentic Systems. The section links point to the parts of the book it draws on.