AI Agents & Tool Use
Models become much more consequential when they can act
What is an AI agent?
“Agent” is not one universally standardized product category. In practice, the term usually refers to an AI system that can pursue a goal across multiple steps by combining a model with some mixture of tools, state, memory, planning logic, and an environment it can affect.
An agentic system can include:
- Model — proposes, interprets, classifies, plans, or generates
- Goal or task — what the system is trying to accomplish
- State — what has happened so far
- Tools — search, code, APIs, files, browsers, databases, messaging, etc.
- Permissions — what the tools are actually allowed to access or change
- Control loop — how the system decides what to do next
- Human controls — approvals, escalation, review
- Monitoring and recovery — logs, limits, rollback, incident response
The model is only one part of the agent.
A simple agent loop
Goal
↓
Observe state
↓
Choose next action
↓
Use tool / respond / ask human
↓
Inspect result
↓
Continue, stop, or recover
Different systems implement this loop differently. Some use explicit planners. Others rely heavily on model outputs. Some use deterministic workflow logic around a model. Some are mostly conventional software with an AI component inside.
Tools change the risk
A model that can only produce text is different from a system that can:
- Execute code
- Browse the web
- Use credentials
- Modify files
- Query private databases
- Send messages
- Spend money
- Change production infrastructure
The important question is not simply “How smart is the model?” It is also what can this system reach, and what can it change?
Memory is a system feature
Agent memory may come from:
- Conversation history
- Saved user preferences
- External databases
- Vector stores
- Files
- Event logs
- Application state
Do not assume every agent has long-term memory, or that memory means the model itself has permanently learned from an interaction.
Planning is not guaranteed
Agentic systems can decompose tasks and adapt to tool results, but they can also:
- Choose the wrong tool
- Misread a tool result
- Repeat actions
- Get trapped in loops
- Follow malicious instructions from untrusted content
- Exceed budgets
- Take actions that are difficult to reverse
“Agent” does not mean “reliable autonomous worker.”
References
Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., et al. (2022). ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023.
Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., et al. (2023). Toolformer: Language Models Can Teach Themselves to Use Tools. arXiv preprint.
Shinn, N., Cassano, F., Gopinath, A., Narasimhan, K., & Yao, S. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. NeurIPS 2023.
Citation Note: Some citations are open access (arXiv or DOI, where linked). Others are books or journal articles that may sit behind a publisher paywall. Use the linked DOI or arXiv when available. If you notice any citation errors, please let us know.