Seven shapes that explain most AI solutions.
An AI solution can be built in a great many ways, and the technical literature catalogues dozens of them. But most of what you will meet in practice can be understood through a small number of recurring shapes. Seven is our choice, not a fact of nature: five workflows you design, and two agents that work more on their own — enough to make sense of almost anything you are shown, and few enough to hold in your head. Each gets a plain definition, a diagram and everyday examples, and the one question that tells you which to reach for: is the path knowable in advance? Written for everyone, from complete beginners to technical specialists.
Is the path knowable in advance?
Every pattern in this guide is a variation on a single question. If you can lay out the steps up front, you have a workflow — steps you design. If you can’t, because it depends on what the AI finds, you have an agent — it works out its own steps to reach a goal you set. Most everyday value, and almost everything worth trying first, sits on the workflow side.
You design the steps
The path is known, so you lay out the steps and the AI runs them. Predictable, easy to check, cheaper. Most business value lives here.
The AI chooses the steps
The path isn’t knowable up front — it depends on what the AI finds — so you set the goal and the limits, and it works out the steps. Flexible, powerful, needs guardrails.
| The question | Workflow | Agent |
|---|---|---|
| Is the path knowable in advance? | Yes — you know the steps. | No — it depends on what the AI finds. |
| Who chooses the next step? | You — you set the steps. | The AI — within limits you set. |
| The trade-off | Predictable, easy to check, cheaper. | Flexible, powerful, needs guardrails. |
The useful question isn’t “is it an agent?” — it’s “how much have I handed over?” The more the AI decides for itself, the more it matters that you set clear limits and stay responsible for the result.
A quick key.
Never read a flowchart before? Every diagram below is built from four simple shapes.
The seven patterns.
Five workflows you design, then two agents that work more on their own. Each gets a plain definition, a diagram, when to reach for it, two everyday examples, how to direct it, and the one thing to watch.
Seven is a working selection, not an exhaustive list. There is no fixed number of AI design patterns, and no official catalogue of them — the open-source list this guide draws on runs to 35, and it keeps growing. These seven are the ones we find most useful for making sense of the AI systems people actually meet at work. Learn them and you will recognise the shape of most solutions put in front of you, including ones that go by a different name. The crosswalk for technical readers further down shows how they map to the fuller list.
Prompt chaining
WorkflowBreak a task into a fixed set of steps, done in order. Each step’s result feeds the next, with a check in between where it helps.
Fail the check → back a step. Put your one check where the work is most likely to go wrong.
the task has clear, ordered steps and the result is better when they are done one at a time rather than all at once.
You design the steps and run them in order — a classic workflow. The skill is clear instructions, and putting your one check at the step most likely to go wrong.
A mistake in an early step carries all the way to the end. Put your check at the weakest link.
Routing
WorkflowSort each incoming request first, then send it to whoever — or whatever — is best placed to handle that type.
the things coming in fall into clearly different types, each needing different handling.
Two parts: a sorter that reads and labels each request, and the specialist paths it feeds. A workflow — you decide the categories.
Always include a clear “doesn’t fit → send to a human” exit. Without one, it will guess when it should have asked.
Parallelisation
WorkflowSplit the independent parts of a task and do them at the same time — then bring the answers together. (Or ask the same question a few times and go with the consensus.)
the parts don’t depend on each other, or you want a second and third opinion to be more confident.
The parts run side by side, then a final step combines them. A workflow. A handy variation: ask the same question a few times and only flag where the answers disagree — a simple reliability check.
Only split parts that are truly independent. If one part needs another’s answer first, that’s step-by-step (Pattern 01), not side-by-side.
Orchestrator and sub-agents
Workflow → AgentA lead AI — the orchestrator — looks at the job, works out what the pieces are, and creates helper AIs (sub-agents) to handle each one, then combines their work. The sub-agents are AI helpers, not people.
Unlike a fixed split, the orchestrator decides how many sub-agents there are, and what each does, at run time.
you can’t list the pieces in advance, because they depend on the particular job in front of you.
You hand the lead AI a goal and let it plan and split the work. This is where a workflow starts to become an agent. Give it a budget and a stopping point — and you still own and sign off the final result.
Left open-ended, it can keep expanding the task. Give it a budget and a clear stop signal.
Evaluator–optimiser
WorkflowOne part produces the work; another checks it against clear standards and sends back feedback. This repeats until it passes.
Doesn’t meet the standard → back to the generator, until it passes.
quality matters and “good” can be written down as a checklist.
Write the standards down and give them to the checker. This is the machine version of a human review — the same skill of checking work before you rely on it.
A vague checker either loops forever or waves everything through. Clear, checkable standards are what make it work.
Autonomous agent (ReAct)
AgentGive the AI a goal and a set of tools; it works in a loop — Reason, Act, Observe — choosing each step itself until the goal is met.
Reason → Act → Observe repeats, the AI choosing each step, until the goal is met.
the path depends on what the AI finds along the way, so you can’t lay out the steps in advance.
You give the goal and let the AI choose the steps. Because it acts on its own, the key is staying in control: set clear limits, keep a human check, and approve anything irreversible before it happens.
Never let an agent take an irreversible action — send, pay, delete, publish — unsupervised. Limit which tools and actions it can use.
Multi-agent — an “agent team”
AgentSeveral AIs, each with a different role, work together as a team — often called an agent team. For example a researcher, a writer and a checker, passing the work between them.
The critic sends the draft back to the writer until it’s approved.
different roles genuinely improve the result and the work can be shared between them.
The most capable shape, but the hardest to keep predictable — more moving parts. You set the roles and stay accountable for what the team produces.
More agents is not better. Add a role only when it genuinely changes the result.
Two things you will hear about constantly are not shapes in their own right. They are options you can switch on inside any of the seven above.
Grounding
often called “RAG”Giving the AI your own documents to work from, so that answers are based on real sources rather than guesswork.
Memory
across steps, or across sessionsLetting the AI carry what it learned from one step, or one conversation, into the next instead of starting from nothing each time.
This is why two systems described as “a RAG chatbot” and “an agent with memory” may in fact be the same shape underneath. Identify the shape first, then ask which capabilities are switched on.
An agent is more like a new colleague than a new tool.
A chatbot is a tool: you operate it, one turn at a time, and nothing happens unless you ask. An agent is different. You give it a goal and it acts on its own — planning, using tools, making choices along the way. That is much closer to handing work to a capable new member of staff than to using a piece of software. And it is the thread running through the seven patterns: as you move from the workflows to the agents, you hand over more of the path.
Think about what you actually do when a capable person joins your team. You don’t give them every system and walk away. You define their role and its limits. You brief them on the goal and the context. You decide what they can sign off alone and what has to come back to you. You review their work before the organisation relies on it. And whatever they produce, the responsibility for it stays with you. Do none of that and a capable hire becomes a liability rather than an asset — decisions no one can explain, work no one checked, mistakes no one caught. Exactly the same is true of a capable agent.
Decide what to hand over
With a colleague: which work is theirs, and which stays with you.
With an agent: which tasks and tools it gets — and which decisions never leave a human.
Brief it well
With a colleague: the goal, the context, and what “good” looks like.
With an agent: the same, written as a brief it can act on — the G-C-O-V method.
Check the work
With a colleague: you review it before the organisation relies on it.
With an agent: the more it does unsupervised, the more this matters — so build the check in.
Stay accountable
With a colleague: their work is your organisation’s responsibility.
With an agent: “the AI did it” is never an answer to a client, a regulator or a colleague.
Write those four down and apply them across the organisation, and they have a name: AI governance.
Governance is simply that management discipline made explicit and organisation-wide: who may use which system, for what, with which limits; who checks the output; who can stop it; and where the record lives. It isn’t a binder you buy or a single person you hire — it’s an operating model, and it pays for itself. It’s the third factor in the AI value equation — the governed organisation that turns capable AI and capable people into value instead of exposure. And it’s the honest answer to the agents your staff may already be running without asking: shadow AI. The wider picture of what can go wrong, and where these controls attach, is the risks of AI, mapped.
One caution, so the comparison doesn’t mislead. Treating an agent like a colleague is about the discipline of managing delegated work — not a claim that the agent is a person. It has no judgement of its own to fall back on and carries no accountability: both stay entirely with you. That is precisely why the limits, the checks and the human sign-off are not optional — and why the most agentic patterns need the most governance around them.
Find your task, find your pattern.
Look at a real task from your own work or life, find the row that fits, and you have your pattern. If more than one fits, pick the simplest — a workflow beats an agent whenever it can do the job.
| Pattern | Reach for it when… | Example | Main skill |
|---|---|---|---|
| Prompt chaining | the task has clear, ordered steps | trip plan; team summary | Clear instructions |
| Routing | requests fall into different types | sorting incoming enquiries | Deciding what goes where |
| Parallelisation | the parts are independent | summarising consultation replies | Clear instructions |
| Orchestrator + sub-agents | the pieces aren’t known up front | impact assessment of options | Handing over + owning it |
| Evaluator–optimiser | quality can be checked against a standard | email to your house style | Checking the output |
| Autonomous agent | the path depends on what it finds | buying a car; research + draft | Staying in control |
| Multi-agent (agent team) | different roles need to collaborate | briefing with a fact-checker | Staying accountable |
The companion worksheet, Map your work to a pattern, turns this into a short exercise in our AI-agents seminars.
How the seven map to the fuller list.
You can skip this. The seven here are a plain-language distillation of a larger, open-source catalogue of 35 patterns — which is itself one catalogue among several, not a closed list. If you want to go deeper, or compare notes with a technical colleague, this is the crosswalk.
| In this guide | Related names in the technical catalogue |
|---|---|
| Prompt chaining | Planning; Self-Discover |
| Routing | Safety & Routing (Dry-Run, Meta-Controller); Adaptive RAG |
| Parallelisation | Self-Consistency, Ensemble; Debate (voting variant) |
| Orchestrator and sub-agents | Planning, PEV; Meta-Controller; STORM |
| Evaluator–optimiser | Reflection, Reflexion, Chain-of-Verification, Constitutional AI |
| Autonomous agent | Tool Use, ReAct, SWE-Agent, BrowserAgent, Computer Use |
| Multi-agent (agent team) | Multi-Agent, Blackboard, Debate, STORM |
| Capability: grounding (RAG) | Agentic / Corrective / Self / Adaptive / GraphRAG |
| Capability: memory | Episodic + Semantic, Graph Memory, MemGPT, Voyager, AWM |
| Advanced / research | Tree of Thoughts, LATS, RLHF Self-Improvement, Cellular Automata |
Where this comes from. The workflow-and-agent distinction follows Anthropic’s Building Effective Agents. The pattern names draw on two open-source catalogues — all-agentic-architectures by Fareed Khan and a companion project by The AI Automators — both freely reusable under the MIT licence. The explanations, diagrams and examples here are Kramer Consulting’s own work, written for a general audience.
From understanding agents to building one.
Our AI-agents seminars run the whole arc — understand what agents are and why they matter, then build your first one in a hands-on day. Small groups, practical from the first hour.
Related guides
From chatbots to agents
AI is moving from answering questions to doing tasks. What’s changing, why it matters, and the skills that stay valuable.
Read the guide Directing agentsDelegate the goal, not the task
A chatbot needs a fresh prompt every time; an agent needs a goal and the room to reach it. The G-C-O-V method for briefing agents — and the real unlock: building the brief into your setup so you stop re-typing it.
Read the guide GovernanceAI governance, explained
What it is, why it pays for itself, and how to make it live in your people instead of a binder nobody reads.
Read the guide