Skip to content

What is an AI agent?

An AI agent is a system that receives an objective, decides which actions are needed to move toward it, can execute those actions through tools, and uses the results to decide the next step. A language model may propose an action, but the complete agent also includes the runtime, tools, state, permissions and the logic that decides when to stop.

The essential distinction is this: a chatbot generates a response; an agent can change the state of another system.

The 60-second answer

An agent is a loop with permissions
The response is not the whole system. The agent observes state, decides an action, executes a tool, checks the result, and decides again.
01Observecontext and state
02Planobjective and next step
03Acttool or environment
04Verifyresult and stopping condition
↺ update state
The practical boundary

Autonomy does not remove control. It moves control into the available tools, permissions, step budget, and stopping policy.

objectivetoolsmemorypolicy

A tool call does not grant authority by itself. The fact that the model emits send_email(...) does not mean the system should execute it. The runtime must validate arguments, permissions, risk and state before producing an external effect.

Chatbot, workflow, copilot and agent

FOUR PATTERNS · WHO DECIDES · WHO EXECUTES
The difference is who chooses the next step and who can execute it
A chatbot, workflow, copilot and bounded agent can all use an LLM; what changes is how much of the path is fixed in code, which decisions are delegated, and where authorization remains.
NOT A RANKINGA deterministic workflow can be the right architecture when the path is already known.
AGENCY SIGNALThe system dynamically chooses which action to use from state and intermediate results.
SAFETY SIGNALChoosing an action and having authority to execute it are different responsibilities.

A workflow is not inferior because it is deterministic. If the steps are already known, it is usually easier to test, explain and constrain. Agency adds value when the sequence depends on the environment and encoding every branch in advance is not worth the complexity.

The components of an agent

A reliable agent needs more than a prompt and several functions:

  1. Objective — what successful completion means.
  2. Context — what information it may use in the current turn.
  3. Tools — which actions exist and what their contracts are.
  4. State — which operations are pending, completed or failed.
  5. Memory — what information may persist between sessions and with what provenance.
  6. Policy — what requires authorization, what is forbidden and what budget exists.
  7. Verification — how the system proves that the result is correct.
AGENT ARCHITECTURE
The model decides; the runtime turns that decision into a controllable system
A reliable agent separates reasoning, authority, state and external effects. The LLM proposes the next step; the runtime keeps the source of truth and applies rules before acting.
1Observable result
2Verify
3Update state
4Next decision

The AI Agents — from responding to acting series develops these components through five progressive chapters.

Tool calling is not the same as agency

An LLM can produce structured arguments for a function. That is tool calling. Agency appears when the system can decide when to use a tool, interpret its result and choose what to do next.

A tool should remain a software contract. The model proposes a call; the runtime keeps the authority boundary and decides whether it may execute.

AUTHORITY BOUNDARY
A tool call is a proposal, not permission
The model can choose an action and construct arguments. Authority to produce an external effect remains in the runtime.
3 · OBSERVABLE RESULT { status, id, error? }
4 · UPDATE STATE pending → running → succeeded / failed
5 · NEXT DECISION continue · retry · respond · stop

Memory, context and state are different things

All three layers participate in the same task, but they answer different questions: context is the model's working view, memory preserves retrievable information, and operational state describes the real execution maintained by the runtime.

THREE LAYERS · THREE RESPONSIBILITIES
Context, memory and state answer different questions
The model receives a working view; memory preserves retrievable information; the runtime maintains the authoritative execution state.
DESIGN RULE Do not derive operational state from what the conversation says. Context = view · memory = persistence · state = execution truth.

The separation becomes critical when a tool takes time. The conversation may record that an action was requested while the runtime knows it is still executing. Operational state — not chat text or retrieved memory — must therefore govern what actually happened and what the system may claim to the user.

How to evaluate an agent

A convincing final answer is not enough. You need to measure the complete task: what the agent decided, what it was allowed to execute, what effect it produced, and whether the system ended in a correct, recoverable state.

AGENT EVALUATION · FULL TRACE
Evaluate the outcome and every transition that produced it
A correct final answer is not enough: verify what the agent decided, what it was allowed to execute, what effect actually occurred, and whether the system remained recoverable.
01 · OUTCOMEDid it reach the objective?Correctness of the complete task, not only the final text.
02 · TRAJECTORYDid it make good decisions?Tool, arguments, order, unnecessary steps, and evidence used.
03 · SAFETYDid it respect the authority boundary?Permissions, approvals, limits, and abstention when authority was missing.
04 · ECONOMYWhat did it cost to get there?Steps, tokens, latency, retries, and risk surface.
05 · RECOVERYWas the resulting state healthy?Idempotency, observable errors, rollback, and safe continuation.
EVALUATION RULE A correct answer does not compensate for a forbidden action, a duplicated effect, or an inconsistent state. The trace turns a failure into a concrete transition that can be reproduced, measured, and fixed.

A trace exposes transitions that a final score hides. It separates a tool-selection failure from an authorization, execution, state, or user-communication failure, and turns each incident into a reproducible case.

How to evaluate an AI agent develops a gate architecture for outcome, trajectory, safety and operational economics.

Why security changes when the system can act

A chatbot that misinterprets a document may produce a wrong answer. An agent with broad permissions can turn the same misinterpretation into an external action.

Security therefore has to live outside the prompt as well: least privilege, per-operation authorization, isolation, human approval for sensitive actions, observability and an external path to stop execution.

The AI Security series covers prompt injection, jailbreaks, poisoned memory, red teaming and production controls.

When using an agent makes sense

Agency is useful when:

  • the objective is clear but the sequence changes depending on what happens;
  • several tools are available;
  • intermediate results determine the next step;
  • the system can verify progress and outcome;
  • permissions and cost can be bounded.

A conventional workflow may be better when the path is known, the error budget is tiny or a deterministic function solves the problem with less risk surface.

Where to go deeper in 5sigmas

Frequently asked questions

Is ChatGPT an AI agent?

It depends on the capability being used. A chat interface that only generates text behaves as a conversational assistant. A system that can choose tools, operate on external resources and continue from their results includes agentic behaviour. The label should describe the real system, not only the model it uses.

Is RAG an agent?

Not necessarily. Retrieving documents and passing them to a model can be a deterministic workflow. It becomes part of an agent when the system can decide when to search, which source to query and what to do next with the result.

Does an agent require multiple models?

No. One model can coordinate multiple tools. Multi-agent systems are one possible architecture, not a requirement.

Does more autonomy mean a better agent?

No. In production, bounded autonomy usually matters more: minimal tools, budgets, verification and a clear stopping criterion.

Primary sources