liveBook · Manning

Guardrail

As explained in Architecting for Autonomy

An architectural control that constrains autonomous system behavior at runtime, independently of the agent's own reasoning.

An agent preparing a contract-dispute response may be allowed to draft internally but not send a response or make a settlement commitment without lawyer approval. A guardrail enforces part of that boundary at runtime—for example, by denying access to the send-response tool. section 5.5.1

Guardrails are architectural controls that constrain autonomous behavior independently of the agent’s reasoning. They hold even when that reasoning is wrong. Unlike a fixed workflow, bounded autonomy permits different paths to an outcome while enforcing limits on what the system can do. section 1.2.2 section 1.6.1

Common runtime controls include:

  • Policy checks that validate actions against authorization rules. section 1.2.2
  • Rate limits that prevent runaway behavior, and hard stops or kill switches that halt or contain execution. section 1.2.2
  • Confidence thresholds and other triggers that escalate decisions beyond safe bounds. section 1.2.2 section 2.2
  • Pre-agent checks that mask sensitive data, detect prompt injection, restrict content, check grounding, or stop tool calls before the model reasons or acts. section 5.5.1

Policy as Data defines the institutional rule: its meaning, owner, applicability, evidence requirements, exceptions, and granted authority. A guardrail applies a selected check derived from that rule; it does not establish the rule’s meaning. A grounding check, for example, can confirm a reference to supplied material but cannot establish that a cited clause governed during a dispute. section 5.5.1

Contract-dispute boundaries

For supervised drafting, pre-agent controls can withhold another matter’s documents, exclude a superseded clause, preserve a transcript’s unverified status, and redact privileged passages from an external draft. The policy record and trusted context explain why an action is prohibited; the tool gateway can enforce the prohibition. section 5.5.1

Instructions in a prompt may shape behavior but do not independently verify compliance: retrieved content can compete with them, tools can provide routes around them, and policy can change. Trace material obligations to their sources, owners, applicability, evaluation results, exceptions, and runtime responses. Enforce them before context, authority, sensitive information, or tool execution crosses a relevant boundary. section 5.7.4

Architects define authority boundaries, escalation thresholds, policy controls, observability, and intervention mechanisms; operations handle events within that design. This runtime governance augments existing infrastructure and processes rather than replacing them. section 2.2 section 1.6.1

How Architecting for Autonomy covers it

How deep it goes
Defines executable control types and explains their relationship to policy, without specifying model-level implementation.
Examples
Uses a contract-dispute drafting scenario to illustrate document filtering, redaction, grounding limits, and denial of external tool access.
What you can do after
Evaluate and understand where independent controls enforce policy boundaries and where checks alone cannot establish policy meaning.
How it connects
Connects guardrails to policy as data, runtime supervision, escalation, and observability.

Also explained in

Compare how 23 books explain guardrail →

Related concepts

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