Map
Framework

Five things to specify for an agent

The question it answers

Before a process is handed to an AI agent, which parts of that agent must its owner write down so the technical team and the business team mean the same thing?

Origin
Anthropic, 2024; Anthropic, 2025; OpenAI, 2025
Diagram type
Hub diagram
Level of analysis
Team
Topics
AI for leadersSupply chain & operations
Glass sphere with a glowing neural-network core, icon for AI for leaders

Diagram

Five things to specify for an agent

Agent specification
Role and instructions
Inputs
Memory
Tools
Boundaries and stopping points

When to use it, and when not to

Use it when

Use it when someone proposes “building an agent” and you need to turn the idea into a description the whole team can read. Fill in the five cells; if one cannot be filled yet, do not build yet.

Avoid it when

The five parts are this site’s grouping of the three sources above. (OpenAI, 2025) names three main components, model, tools and instructions; splitting out role, inputs, memory and boundaries is the site’s way of dividing them. If a single model call with retrieval is enough, do not build an agent: (Anthropic, 2024) advises choosing the simplest solution.

How to apply it, step by step

  1. Role and instructions: what the agent does and by which procedure. (OpenAI, 2025) advises building on existing operating procedures, support scripts or policy documents, and making each step correspond to a specific action.
  2. Inputs: the agent starts from a command or a conversation with the user ((Anthropic, 2024)).
  3. Memory: what the agent keeps between steps. (Anthropic, 2024) lists memory, along with retrieval and tools, as an augmentation of the model; (Anthropic, 2025) describes structured notes stored outside the context window.
  4. Tools: the data the agent can read and the actions it can take. (OpenAI, 2025) sorts tools into data, action and orchestration; (Anthropic, 2025) warns that bloated, overlapping tool sets make it hard for the agent to choose.
  5. Boundaries and stopping points: when the agent stops and when it hands back to a person. (Anthropic, 2024) names stopping conditions such as a maximum number of iterations and pauses for human feedback; (OpenAI, 2025) places guardrails around the agent.

Pitfalls

Pitfall

Writing “the agent decides” without writing the stopping points. (Anthropic, 2024) says agents cost more and can compound errors, so they need testing in a sandbox with suitable guardrails.

Sources

References

Listed A–Z by author surname, Harvard style. A source with a DOI links to https://doi.org/…; one without links to its page or says plainly that it has no DOI.

  1. Anthropic (2024) Building effective AI agents. Anthropic Research.No DOIanthropic.comCited in 6 articles
  2. Anthropic (2025) Effective context engineering for AI agents. Anthropic Research.No DOIanthropic.comCited in 4 articles
  3. OpenAI (2025) A practical guide to building agents [guide].No DOIcdn.openai.comCited in 3 articles