Agentic Patterns

Effective agentic systems are usually built from a small set of composable patterns rather than one monolithic "do everything" agent. ByteChef lets you build each of these patterns on the workflow canvas by combining flow controls with the AI Agent — no framework code required.

The five patterns below follow the taxonomy from Anthropic's Building Effective Agents. Each maps onto ByteChef primitives you already have.

---

Prompt Chaining

Decompose a task into a fixed sequence of steps, where each step's output feeds the next. In ByteChef this is simply a linear sequence of tasks on the canvas — one or more AI Agent (or LLM) steps in order, passing data with data pills. Add a Condition between steps as a gate when a stage should only run if the previous output passes a check.

Routing

Classify an input and send it down a specialized path. Use a Condition or Branch flow control — often driven by a first AI Agent step that classifies the input — to route each case to the handler best suited to it. Each branch can use a different model, prompt, or set of tools.

Parallelization

Run independent subtasks at the same time, then combine the results. ByteChef offers several fan-out flow controls depending on the shape of the work:

• Each / Map — apply the same steps to every item of a list in parallel (Map collects the results).
• Parallel — run several distinct task lists concurrently.
• Fork-Join — run isolated branches in parallel, each with its own context, then join.

Orchestrator-Workers

A central agent breaks a complex job into subtasks and delegates each to a specialized worker. Build this inside a single agent with the Task tool from the Agent Utils toolset — the agent dynamically delegates to specialized sub-agents at runtime. Or build it as a workflow topology with Subflow, where an orchestrating workflow starts child workflows that run in isolation and report back.

Evaluator-Optimizer

One agent produces a result; another evaluates it and feeds back improvements, looping until the result is good enough. Combine a Loop flow control with two AI Agent steps — a generator and an evaluator — using the evaluator's verdict as the loop's continue/break condition. For grading agent outputs systematically, see Agent Evals.

---

Combining Patterns

Real systems compose these patterns: a router that dispatches to parallelized workers, an orchestrator whose workers each run their own chain, an evaluator-optimizer loop wrapped around any of the above. Because every pattern is built from the same flow controls and agent nodes, they nest and combine freely on one canvas.