Primer
Tools, Agents, and Orchestration
Choose among a model call, typed tool, bounded agent loop, explicit graph, and multi-agent design without adding unnecessary machinery.
- 01Distinguish a model call, typed tool, bounded agent loop, explicit graph, and multi-agent pattern
- 02Explain tool, skill, plugin, and MCP at the correct abstraction layer
- 03Choose the simplest orchestration structure that meets the task's control needs
Five control structures
architecture map| Structure | What controls the work |
|---|---|
| Model call | One request produces one response |
| Typed tool | Code exposes a named operation with an input and output contract |
| Bounded agent loop | A model chooses actions repeatedly within budgets and stop conditions |
| Explicit graph | Nodes and edges define branches, joins, retries, waits, and human gates |
| Multi-agent pattern | Separate agent contexts or policies coordinate through an explicit workflow |
One request can gain typed operations, bounded iteration, explicit graph structure, and—only when justified—separate agent boundaries.
Text description
One request can gain typed operations, bounded iteration, explicit graph structure, and—only when justified—separate agent boundaries.
Tools connect generation to action
A tool is a typed operation such as search_orders, create_ticket, or run_test. The model may
propose a tool name and arguments, but the host validates authorization and inputs, executes the
operation, and returns the result. A tool contract narrows communication; it does not make the
model trustworthy or the operation safe by itself.
A bounded agent loop repeats a cycle such as model → tool → result → model. It needs maximum steps, time, token or cost budgets, no-progress detection, validation, and explicit success and failure exits.
An explicit graph is useful when dependencies matter: two branches can run in parallel, a join waits for both, one failure retries locally, or a human must approve a write. A graph can use one model or no model in some nodes. It does not imply multiple agents.
| Choose | When it is the simplest fit |
|---|---|
| One call | The answer can be produced from available context without external action. |
| Tool use | The model needs a narrow operation or fresh external data. |
| Bounded loop | The next action depends on the previous result and may repeat. |
| Explicit graph | Branches, joins, retries, waiting, or ownership must be visible. |
| Multiple agents | Separate boundaries provide real isolation, specialization, parallelism, or ownership. |
Tool, skill, plugin, and MCP
- A tool is one executable, typed operation.
- A skill is a reusable capability made from instructions, examples, workflows, or tools.
- A plugin is packaging installed into a particular host product.
- MCP is an interoperability protocol through which a host uses clients to connect to servers that expose capabilities such as tools, resources, and prompts.
These terms are related but not interchangeable. The dedicated capability module covers MCP transport, lifecycle, and trust boundaries in detail.
Common MisconceptionA serious agent system should have several agents.
CorrectionMultiple agents add handoffs, duplicated context, latency, failure modes, and evaluation work. Add them only when a concrete boundary pays for that cost.
Deep dive: Multi-agent is a boundary decision
Separate agents can protect a small context from unrelated work, give a specialist a restricted tool set, parallelize independent tasks, or map responsibility to different services or teams. If the agents share the same context, tools, policy, and owner and run sequentially, separate names may add ceremony without adding a useful boundary.
Orchestration mistakes
- Giving a write tool to a model without authorization and argument validation in the host.
- Calling repeated model-tool execution an agent without budgets or stop conditions.
- Using a graph diagram that is not the control structure the runtime executes.
- Splitting one coherent task across agents and paying coordination cost without a clear boundary.
References & deeper reading (3)
- Building Effective AgentsAnthropic Engineering
- MCP ArchitectureModel Context Protocol
- LLM FundamentalsMicrosoft Learn
Retrieval Practice
Check one idea at a time
Order these structures from the smallest control surface to the largest.
Put these in the correct order.