AI agents
AI agent
An AI agent is software in which a language model decides its own next steps and tool calls in a loop, using results from its environment, until a goal is met or a stopping condition is reached.
What it is
Anthropic's engineering guide draws a useful line. Workflows are systems where models and tools are orchestrated through predefined code paths. Agents are systems where the model dynamically directs its own process and tool use, keeping control over how it accomplishes the task. Both start from what the guide calls the augmented LLM: a model given retrieval, tools and memory.
Why it matters
An agent can take on open-ended work no one scripted in advance: triage a queue, reconcile two systems, chase an overdue account. The same freedom is the risk. Each step builds on the last, so one wrong read or one wrong call can carry through the rest of the run, and an agent with write access acts in real systems at machine speed.
How it works
- The agent receives a goal and a set of tools.
- It plans a step and calls a tool.
- It reads the result from the environment, the ground truth for its next decision.
- It repeats until the goal is met, a person is needed, or a stopping condition such as a maximum number of iterations ends the loop.
Anthropic's guide recommends checking real results at every step, pausing for human input at checkpoints, extensive testing in sandboxed environments and appropriate guardrails.
Example
A service-desk agent is asked to find clients who are both overdue on invoices and have an open urgent ticket. It lists overdue invoices in the accounting system, lists urgent tickets in the PSA, matches the two by customer, and drafts a note for the account manager. It does not send the note; a person does.
Common pitfalls
The OWASP Top 10 for LLM applications names the failure excessive agency, with three root causes:
- Excessive functionality. The agent can reach tools it does not need.
- Excessive permissions. Its credentials allow more than the task requires, such as write access for a read-only job.
- Excessive autonomy. High-impact actions run without independent confirmation.
Add to these: credentials pasted into prompts, loops with no budget, and no record of what the agent did or why.
How actionlake handles it
Through actionlake, an agent reaches every connected system from one MCP endpoint and stays autonomous only within the limits you set. Each agent gets its own key with a scope (read-only, one service, one action, or more), a spending cap refused before dispatch, and approval rules for the actions that should wait for a person. The agent never holds a vendor credential. Every do and ask carries a one-sentence intent, and every run goes into the ledger. Revoking a key turns that agent off, along with every key it created. See how teams use it.
Questions
What is the difference between an AI agent and a chatbot?
A chatbot answers in text. An agent uses tools in a loop to take steps toward a goal, such as reading records and creating tickets, deciding each next step from the results.
What is the difference between an AI agent and a workflow?
In Anthropic's terms, a workflow follows predefined code paths that call a model at set points, while an agent decides its own steps and tool use as it goes.
How do you stop an AI agent doing damage?
Limit what its credential can reach, cap what it can spend, require a person's approval for high-impact actions, and record every action with its reason. Enforce all four outside the model.
Related terms
- Agentic workflowAn agentic workflow is a multi-step process in which a language model plans or performs some of the steps with tools, ranging from fixed code paths that call a model at set points to open-ended agents that choose every step.
- Tool callingTool calling, also called function calling, is a language model's ability to answer with a structured request to run a named function with arguments, which the application executes before handing the result back to the model.
- AI agent permissionsAI agent permissions are the rules that decide which systems, actions and data an agent's credential may reach, and least privilege for agents means granting only the actions a task needs and enforcing that limit outside the model.
- Human-in-the-loop approvalHuman-in-the-loop approval is a control that pauses an AI agent's action until a named person approves or rejects it, so high-impact or irreversible actions never run on the model's judgement alone.
- Agent spending capAn agent spending cap is a hard limit on how much a single AI agent credential may spend or call in a period, checked before each request is sent, so that a looping or misbehaving agent stops at the limit instead of after it.
- AI agent audit logAn AI agent audit log is a record of every action an agent took: which credential ran it, what it did, in which system, with what outcome and cost, and why, kept so a person can reconstruct and verify events later.
Sources
Connect an agent. Start free.
100,000 calls free every month. No card. Hard caps on every key.