How we build

Choose the simplest design that works, then earn the complexity.

The model is one part of an AI system. The structure around it, how steps are ordered, checked and handed to people, decides whether it is reliable and affordable.

Book a 30-minute call →See the six patterns

A hand drawing a simple flowchart on paper

Six designs, from a fixed path to a free-running agent

Each step up gives the model more freedom and costs more to run and to trust. We climb only as far as the problem needs.

Prompt chaining

The task is split into a fixed sequence. Each step takes the previous step's output, and a check between steps catches a bad result before it travels on.

Good fit: The work breaks cleanly into known steps and you would rather be accurate than instant.
Example, illustrative: Turn a refund policy into a customer reply: pull the facts from the order, draft the reply, then check tone and policy before it is sent.
Trade-off: Slower than one call, and every added step adds cost.

Routing

A first step classifies the incoming request and hands it to a prompt, tool or model built for that kind of request.

Good fit: There are clearly different kinds of input, each handled better on its own, and sorting them is reliable.
Example, illustrative: Support mail arrives as billing, bug or how-to. Each goes to its own handler, and easy questions can use a cheaper model.
Trade-off: A wrong sort sends the request down the wrong path, so the sorter needs its own tests.

Parallelization

Work runs side by side and the results are combined. Either independent parts run at once, or the same task runs several times and the answers are compared.

Good fit: The parts really are independent and speed matters, or one answer is not enough and you want several views.
Example, illustrative: Review a contract for several separate risks at the same time, or have three reviewers flag the same clause and escalate when they disagree.
Trade-off: You spend more to save time or buy confidence.

Orchestrator and workers

A central model decides at run time how to split the task, hands the pieces to worker models and merges what comes back.

Good fit: You cannot list the steps in advance, as in research across many sources or a change that touches many files.
Example, illustrative: Research a market: the lead decides which questions matter, workers read different sources, and the lead writes the combined brief.
Trade-off: The most flexible of the fixed designs and the least predictable, and it costs the most.

Evaluator and optimizer

A generator produces an answer. An evaluator scores it against written criteria and sends feedback until it passes or the limit is reached.

Good fit: You can say clearly what good looks like, and a second look measurably improves the result.
Example, illustrative: Write product descriptions, then score each against the brand guide and the facts in the catalogue, and rewrite until it passes.
Trade-off: Only works when the criteria are clear. Vague criteria give a loop that polishes the wrong thing.

Agents

The model decides its next step, calls tools, reads what happened and repeats, pausing for a person at agreed checkpoints or when it is stuck.

Good fit: The problem is open-ended, the number of steps cannot be fixed in advance, and the environment is safe enough to give it room.
Example, illustrative: Work through a support case across the helpdesk, the order system and the knowledge base, and ask a human before issuing a refund.
Trade-off: Higher cost and higher risk than a fixed design, so it needs sandbox testing, limits and approvals.

How we choose

Start from your situation. The first row is where we begin on every project.

Your situationStart withIt costs you

One question, one answerA single call with contextLeast cost and delay

Steps are known and orderedPrompt chainingSlower, more calls

Different kinds of requestRoutingNeeds a reliable sorter

Independent parts, or need consensusParallelizationMore tokens for speed or confidence

Steps unknown until you startOrchestrator and workersLeast predictable

Clear quality bar, drafts improve on reviewEvaluator and optimizerNeeds written criteria

Open-ended, tool-heavy, long-runningAgentsMost cost and risk, so approvals matter

Four rules we build by

The same principles apply whichever pattern we use.

01
Simplest first
A single well-prompted call with the right context is often enough. We add structure only when it measurably improves the result.
02
Visible steps
Plans, tool calls and decisions are written to a log you can read, so you can see why the system did what it did.
03
Careful tools
A tool is an interface the model has to use correctly. We write clear descriptions, give examples and design them so mistakes are hard to make.
04
Tested before trusted
We agree success criteria first and test against real examples before launch, and we keep a person in the loop where an error is costly.

Questions about our approach

Do we always need an agent?

No. Many jobs are better served by a fixed sequence or a single well-prompted call, which are cheaper, faster and easier to test. We recommend an agent only where the steps genuinely cannot be fixed in advance.

How do you decide which pattern to use?

We start from the simplest design that could work, test it against real examples, and add structure only where the results measurably improve. The AI readiness audit does this ranking for you.

Can we see what the system did and why?

Yes. Plans, tool calls and decisions are logged in a form your team can read, and risky actions wait for a person to approve them.

Where do these patterns come from?

The pattern names follow Anthropic's published guide, Building effective agents. How we apply them, test them and hand them over is Austial's own work.

Pattern names follow Anthropic's published guide, Building effective agents. Austial is an independent studio and is not affiliated with Anthropic. Examples above describe the shape of the work and are illustrative, not client results.

Not sure which pattern fits?

Tell us the job. We will tell you the simplest design that can do it, in a 30-minute call.

Book a 30-minute call →Send a project brief