Corpus Agentis
The field book to agent ecosystems
The field book to agent ecosystems
Overview · The distinction

What is an agent?

A model generates. An agent acts. A model is a stateless reasoning engine: information goes in, an answer comes out. An agent is the runtime around that intelligence, combining models with memory, tools, control flows and governance to pursue a goal over time.

The containment
Model Agent GPT · Claude Gemini · Llama A stateless function. Text in, text out. wrapped in Model Transports Effectors Memory Compute Orchestration Governance State Perimeter A runtime loop with memory, tools and control flow. Uses one or more models as its reasoning core.
Comparing models and the agents that run on them
Dimension
Model
Agent · what the Fieldbook maps
Reasoning
The trained network is the reasoning engine.
Uses the model as its reasoning core.
Memory
None. Stateless between calls.
Session, task and durable planes.
State
None. Every call is independent.
Carried through the loop; survives crashes and approval waits.
Orchestration
None. Single-turn function call.
The control loop, classify · confirm · dispatch.
Tools · Effectors
None. Can only produce text.
Tool calls, MCP connectors, file writes, API actions, gated by tier.
Transports
N/A. Text in, text out.
Chat, voice, API, MCP, agent-to-agent, computer-use.
Skills
No skill sockets. Just what the weights know.
Task-specific behaviours bound to skill sockets. Composable.
Compute
An accelerator running inference. Cost per token.
Same, plus the runtime it deploys on.
Governance
No perimeter, no policy.
Sandboxes, auth, approval gates, capability tiers.
Autonomy
Responds only when called.
Can act between turns, run background loops, hand off to other agents.
Examples
Thing
Is a
Why
GPT-5 · GPT-4o (API)
MODEL
The weights and inference endpoint OpenAI sells. Text in, text out. No memory of you between calls.
Claude Opus 4.7 · Sonnet 5 (API)
MODEL
Anthropic's underlying models. Same shape, stateless function accessed through an API.
Gemini 2.5 Pro · Flash (API)
MODEL
Google's foundation models. What Vertex AI serves under the hood.
ChatGPT (web / iOS / desktop)
AGENT
Uses GPT models. Adds conversation memory, tools (browse, code interpreter, image gen), file uploads, custom instructions.
Claude.ai · Claude Desktop
AGENT
Uses Claude models. Adds Projects for context, tools via MCP, artifacts, Computer Use.
Cursor · Windsurf · GitHub Copilot Chat
AGENT
Uses multiple models. Adds a code-focused loop: file edits, terminal, git, project index, review-and-apply flow.
Microsoft Copilot (M365 · Office · Teams)
AGENT
Uses OpenAI's GPT models via Azure. Adds enterprise data grounding (Microsoft Graph), tool routing into Office actions, per-tenant memory and a governance perimeter.
Devin · Cognition AI
AGENT
Autonomous software engineer. Long-horizon loop with a sandboxed VM, browser, editor, and its own planning stack.
ChatGPT · Claude Agent Mode
AGENT
Same underlying model, expanded runtime, multi-step task execution, tool use, self-verification.
openai.chat.completions.create({...})
MODEL
A single API call. Returns text and forgets you existed. Not an agent; a raw model invocation.
Field notes

Brain vs body, actions vs understanding.

An agent uses one or more models as its reasoning core, sometimes as a team of specialist sub-agents, and wraps them in the systems required to operate: memory, transports, effectors, orchestration and guardrails. That runtime keeps state between turns, acts in external systems, inspects what happened and decides what to do next. Think of the model as the intelligence and the agent as the operating system that puts that intelligence to work.

From the corpus, curated by Brandon Chaplin
Common questions
Isn't ChatGPT a model?

No. ChatGPT is a product, a web / desktop / mobile agent, that runs on top of OpenAI's GPT models. GPT-5 and GPT-4o are the models. ChatGPT is the runtime wrapping them: conversation memory, browsing, code interpreter, file uploads, image generation, custom instructions. Same distinction for Claude the product (agent) vs Claude Sonnet 5 the model, and Gemini the assistant (agent) vs Gemini 2.5 Pro the model.

Where is the boundary drawn?

State and action. A model has no memory of the last call and cannot take actions. An agent has memory (short-term and durable) and can call tools that touch the world. The moment you add a loop that maintains context and can execute a tool, you have an agent, even a very simple one.

Why does this distinction matter?

Benchmarks measure models, real deployments measure agents. When a paper says "GPT-5 achieves 80% on X," the same model inside a different agent stack, different memory, different tools, different orchestration, will produce different numbers. Every "AI beats human" figure in the productivity literature is really an agent-with-a-model figure. The corpus you're reading maps agents; the reasoning core inside them is the model.

Are there hybrid cases?

A raw API call is a model invocation. A single-turn chat with no tools is barely an agent, technically a runtime, but the anatomy is minimal. The interesting agents are the ones that maintain state across turns, invoke tools under a control loop, and enforce a governance boundary on what they can and cannot do. See Agent Anatomy for the parts.

Next in the learning path