Agentic AI Patterns
From single agents to orchestrated teams — the architecture patterns behind AI that acts, and how to choose between them without over-engineering.
The patterns
Click each pattern to see the architecture and business context.
The building block
Single Agent
"One employee with clear instructions and access to tools"
A single agent receives a goal, breaks it into steps, and uses tools to accomplish them — one action at a time. It reads data, calls APIs, writes files, and checks its own work. The key difference from a chatbot: it doesn't just answer questions, it takes actions. It has a loop: think, act, observe, repeat.
Start here. Most agent use cases don't need multi-agent orchestration. A single well-prompted agent with the right tools handles 80% of automation tasks. Over-engineering with multiple agents when one would do is the most common mistake in agentic AI.
The building block
Single Agent
"One employee with clear instructions and access to tools"
A single agent receives a goal, breaks it into steps, and uses tools to accomplish them — one action at a time. It reads data, calls APIs, writes files, and checks its own work. The key difference from a chatbot: it doesn't just answer questions, it takes actions. It has a loop: think, act, observe, repeat.
Start here. Most agent use cases don't need multi-agent orchestration. A single well-prompted agent with the right tools handles 80% of automation tasks. Over-engineering with multiple agents when one would do is the most common mistake in agentic AI.
The team
Multi-Agent Orchestration
"A project manager delegating to specialists"
An orchestrator agent receives the goal and delegates sub-tasks to specialist agents — a researcher, a writer, a reviewer, a coder. Each specialist has its own system prompt, tools, and context window. The orchestrator coordinates, merges results, and handles failures. Like a real team: the manager doesn't do the work, they make sure the right people do.
Use multi-agent when the task genuinely requires different expertise or when context windows aren't large enough for a single agent. The coordination cost is real — message passing, error handling, state management. Don't split into agents what one agent can handle sequentially.
The safety net
Human-in-the-Loop
"A junior employee who escalates to their manager for big decisions"
The agent works autonomously for routine tasks but pauses at defined checkpoints: before sending an email, before executing a payment, before deleting data. A human reviews, approves or rejects, and the agent continues. The boundaries are explicit: below this threshold, act freely; above it, ask.
This isn't a limitation — it's a feature. The highest-value agent deployments all have human checkpoints for irreversible or high-stakes actions. The goal isn't full autonomy; it's appropriate autonomy. Start with tight guardrails and widen them as trust builds.
The oversight layer
Human-on-the-Loop
"A factory supervisor watching the production line from the control room"
Unlike human-in-the-loop (where the agent pauses and waits for approval), human-on-the-loop means the agent acts autonomously while a human monitors the output. The human can intervene, correct, or shut things down — but they don't block each action. Think of it as supervision rather than co-signing. The agent sends periodic summaries, flags anomalies, and the human reviews asynchronously.
This is where mature agent deployments land. You start with human-in-the-loop (tight guardrails), measure error rates, and gradually shift to human-on-the-loop as confidence builds. The economics are better: the human reviews a dashboard of 50 completed actions rather than approving each one individually. Reserve in-the-loop for truly irreversible actions.
Universal connectivity
Tool Use via MCP
"USB-C for AI agents — connect once, use everything"
An agent without tools is just a chatbot. Tools let agents read databases, call APIs, search the web, write files, and interact with any system. MCP (Model Context Protocol) standardises this: instead of building a custom integration for every tool, you expose tools via a single protocol. Any MCP-compatible agent can use any MCP-compatible tool.
MCP eliminates the N×M integration problem. Without it, 5 agents × 4 tools = 20 custom integrations. With MCP, it's 5 + 4 = 9 implementations. The protocol is the multiplier that makes agentic AI practical at enterprise scale.
Revenue recovery agent
Revenue Recovery Agent
"An accounts receivable clerk who never sleeps"
Here's a concrete example that ties all the patterns together. A revenue recovery agent monitors overdue invoices, checks the CRM for context (are they a key account? is there a dispute?), drafts an appropriate follow-up email, escalates to a human for high-value accounts, and logs every action. It runs on a schedule, handles edge cases, and gets better with feedback.
This single agent combines: tool use (CRM, email, database), human-in-the-loop (escalation for high-value accounts), and a clear success metric (recovered revenue). It's not a demo — it's a pattern that applies to any process where humans currently chase status and send follow-ups.
Open the explorer Agentic Patterns Explorer
Compare canonical agentic patterns side-by-side — trade-offs, best-fit workloads and worked examples, in an interactive pattern explorer.
Decision framework
I'm speaking on this — The Compute Infrastructure Questions Every AI Buyer Should Ask →