Agent
An agent is a model using tools in a loop until the job is done. In practice: Agent = Model + Harness. The model reasons; the harness is everything else that makes it safe and useful.
1. Chatbot vs agent
Chatbot
One prompt in, one answer out. Can only say things.
Agent
Acts, looks at the result, decides the next step. Can do things and check them.
2. The agent loop
Every turn: assemble context → model picks an action → harness runs it → result goes back into context. It stops when the model answers without asking for a tool (or a limit is hit). Anthropic's short version: gather context, take action, verify work, repeat.
Context window, right now
Everything the model knows this turn. It only grows.
3. The harness
The harness is the runtime around the model. It is not the model, but it decides what the model sees, what it may do, and how it knows it is right. Five layers (click one):
Example: this lab's own AGENT.md (see How the Lab works) is harness context: a file the
agent reads every session so it knows the project's goals and rules.
4. Four “engineerings”, nested
Each one wraps the previous. Smaller scope inside, longer time horizon outside.
| Layer | Question it answers |
|---|---|
| Prompt | How do I word this one request? |
| Context | What is the smallest set of useful information to show the model this turn? |
| Harness | What tools, checks, permissions and memory surround a whole run? |
| Loop | How do repeated runs proceed: budgets, retries, when to stop, when to ask a human? |
5. How it happened
Each step fixed the previous step's biggest failure.
- Oct 2022 · ReAct Reason and act interleaved in a loop.Problem: models could only talk. Fix: let them call tools and read results.
- Mar–Jun 2023 · AutoGPT, function calling Open-ended loops; tool calls become a first-class API feature.Problem: parsing tool requests out of free text was fragile. Fix: structured tool calls.
- Nov 2024 · MCP A standard way to plug tools into agents.Problem: every agent re-wired every tool. Fix: one protocol.
- Dec 2024 · “Building effective agents” Anthropic: agents are LLMs using tools in a loop; start simple, add complexity only when needed.Problem: over-built frameworks. Fix: simple, composable patterns.
- 2025 · Coding agents, context engineering Agents run for dozens of steps and the context fills up and drifts.Problem: long runs degrade. Fix: curate what is in the window: retrieve, summarise, isolate, offload.
- 2026 · Harness and loop engineering The model is good enough that results depend on what surrounds it, and on running it repeatedly with limits.Problem: reliability and trust. Fix: verification, guardrails, traces, budgets, retries, escalation.
6. Takeaways
- An agent is a loop: context → model → tool → result → context. No tools means a chatbot.
- Quality comes less from the model's words and more from what it sees and how it is checked.
- Verification (tests, linters, real output) is the loop's reward signal. Prefer deterministic checks.
- Start with the simplest loop. Add harness layers only when a failure demands one.
Sources
- Anthropic, Building effective agents (agent definition, ground truth from tool results)
- Anthropic, Effective context engineering for AI agents
- Oracle, What is the AI agent loop?
- IBM, What is loop engineering?
- AI Harness Engineering: A Runtime Substrate for Foundation-Model Software Agents (arXiv 2605.13357), five-layer harness framing via Faros
- “Agent = Model + Harness” attributed to Mitchell Hashimoto by Atlan; not checked against the original.