01
Automation
Executes a predefined path.
No interpretation, no choice between actions. A full component of a system, and frequently the right one.
Agent anatomy / 002
Different parts.
Explicit roles.
Five terms do different jobs, and conflating them is how the word agent stops meaning anything. Here is what each one is, and the test that separates it from the next.
A shared vocabulary
Not every automation
is an agent.
Five terms, one job each. If a thing never interprets context, makes a bounded decision or routes an exception, it is automation — a full component of a system, and often the better engineering result. The distinction exists so the right mechanism gets picked, not so agency ranks above the rest.
Our own registry labels each entry accordingly, including the ones that are not agents.
02
Agent
A bounded responsibility that chooses, recommends or routes a permitted action from context.
The decision is the point. It can be entirely deterministic — an LLM is not required.
03
LLM
An optional component for interpreting or generating language where ambiguity matters.
Neither agency nor authority by itself.
04
Harness
The technical layer giving an agent controlled access to context, tools and state.
It enforces policy, permissions, logging and escalation. Infrastructure, not the outcome.
05
System
Combines automations, agents, humans, rules, tools, data and controls to deliver an outcome.
The outcome belongs to the system, never to one component in it.
What makes an agent inspectable
01
Context
What can the system perceive and remember?
Observations, instructions, knowledge, state, history and retrieved context give a system an operational view of its environment.
- Observations
- Knowledge
- Memory
- State
- History
02
Capabilities
What can the system determine and do?
Capabilities combine the appropriate mechanisms—rules, algorithms, models, tools and communication—to perform a defined role.
- Classify
- Reason
- Decide
- Plan
- Execute
03
Activation
When does the system operate?
A persistent system behaves differently from a prompt-response tool. Its activation must be explicit.
- Event
- Schedule
- Threshold
- Human
- System call
04
Authority
What is the system allowed to change?
Permissions, limits, approvals and escalation define the difference between a recommendation and an action.
- Read
- Write
- Limits
- Approval
- Escalation
05
Verification
How do we know the system behaved correctly?
Evaluation, traces, outcomes and feedback make consequential systems inspectable over time.
- Evaluation
- Logs
- Outcomes
- Feedback
- Policy checks
Composition
The configuration changes with the operating problem.
Rules can be safer than model reasoning. Human judgment can be essential. The practical question is which combination can reliably perform the required behaviour.
Inspect a system specification ↗