In brief
- An agent combines a model with instructions, context, tools and an execution loop.
- Tools may search, calculate, create documents or trigger actions in business systems.
- The broader the freedom to act, the more important permissions, approvals, logs and stop conditions become.
What is an agent made of?
The agent is not only the language model. The application defines the objective and rules, supplies context and exposes tools. An execution mechanism passes intermediate results back to the model until the goal is reached, a limit is triggered or a human decision is required.
- Objective and boundaries
- Define the required outcome, permitted subtasks and explicit stop or escalation conditions.
- Model
- Interprets the current state and proposes a next step. It is not accountable and must not assign permissions to itself.
- Context and memory
- Hold the assignment, intermediate results and approved information for the current case. Long-term retention is a separate decision that must be designed deliberately.
- Tools
- Enable retrieval, calculations or actions. Every tool needs a narrow purpose, validated parameters and minimal technical permissions.
- Execution and control
- Software checks permissions, limits steps, time and cost, logs calls and requires approval for consequential actions.
How does an agent work?
A typical cycle starts by interpreting the assignment. The agent plans a next step, calls a tool, evaluates its result and decides what to do next. This can adapt flexibly to different cases, but it is less predictable than a fixed process.
- Step 1
Understand
Interpret the assignment, context and boundaries.
- Step 2
Plan
Select a suitable next step or sequence.
- Step 3
Act
Call an approved tool with limited parameters.
- Step 4
Evaluate
Assess the result, continue, correct or escalate.
When is an agent useful?
Agents suit variable tasks where the route and required information are not always the same. Examples include research across several sources, preparing complex cases or coordinating different tools. The objective must still be clear and the result reviewable.
If every step is known and stable, a normal workflow is usually easier to test and operate. An agent should add flexibility where it creates real value, not become the default for every automation.
Which controls does an agent need?
An agent can multiply an error if it performs several actions from a false assumption. Tools that write, communicate externally or cause irreversible change require especially tight constraints.
- Expose only necessary tools and minimum permissions
- Limit amounts, recipients, step count and runtime
- Require confirmation before critical actions
- Log every tool call and result
- Provide safe stop, rollback and escalation paths
How is an agent evaluated?
Assess the path as well as the outcome: Did the agent choose the right tools, avoid unnecessary steps, respect permissions and stop under uncertainty? Evaluations should cover normal cases, exceptions, failing tools and manipulated inputs.
Example: preparing a supplier case
A new supplier requires document checks, requests for missing evidence and preparation of data across several systems. An agent reads the case status, selects the relevant checklist, looks up validity dates and drafts a request. It receives read access and a few preparatory tools; an accountable person reviews and approves the setup. The agent cannot change bank details or give final supplier approval.
What to remember
Give an agent only the freedom the variable task genuinely needs. A tightly bounded agent with explicit approval is often more valuable than a system with maximum freedom to act.
Sources and further reading
These primary sources provide further detail on definitions, technical foundations or responsible use.
Content reviewed