Engineering · Systems Engineering

Why Reliable Agents Need More Than a Model

The surrounding infrastructure software workers need: permissions, tools, state, policies, testing, approvals, exceptions, and observability.

Author
Quaspar Research
Published
September 21, 2026
Last reviewed
Last reviewed September 21, 2026
Reading time
11 min read
Article contents

Intelligence is one dependency

Operational software must know who it represents, which systems it may access, which actions it may perform, what state it owns, and when it must stop for review.

The quality of generated text does not answer those questions. They belong in explicit infrastructure that can be tested, observed, and changed independently.

The surrounding control plane

  1. 01Identity and scoped credentials
  2. 02Typed tools and system connections
  3. 03Durable workflow state
  4. 04Policies and action boundaries
  5. 05Simulation and evaluation
  6. 06Approval and escalation
  7. 07Execution evidence
  8. 08Monitoring, rollback, and recovery

Separate concerns deliberately

Model

Interprets context or proposes a next step within an assigned task.

Policy

Determines whether a proposed action is allowed, reviewable, or blocked.

Tool

Executes a narrowly defined operation against a system.

Workflow

Carries state, retries, exceptions, approvals, and completion criteria.

Evidence

Makes the run reconstructable without relying on a conversational transcript.

Reliability is a system property

Reliable behavior comes from the interaction of bounded components, not from a prompt asking a model to be careful. The architecture should make unsafe or ambiguous actions difficult, visible, and interruptible.

Sources

Authoritative references

Rules, standards, and guidance can change. The links below were reviewed on September 21, 2026.

  1. 1Assistant Secretary for Technology Policy / Office of the National Coordinator. ONC Conformance Test Tools