Overview

Agent work gets hard to reason about when every run is described as a loose sequence of prompts. The useful shape is usually smaller: a chain, a branch, a gate, a loop, or a reduction over several attempts.

This pattern language gives those shapes names. It is meant for planning and reviewing agent systems, not for drawing architecture theater. Each pattern should make the control flow obvious: what starts, what branches, what must be true before continuing, and what evidence decides whether the work moves forward.

Primitives

These are the primitives I reach for first. Larger agent workflows are usually compositions of these shapes, with domain-specific details attached at the edges.

One step finishes before the next begins.

Use when ordering matters: setup before implementation, tests before publish, or a migration before a dashboard update.

One intent creates multiple independent branches of work.

Use for delegation, research shards, package splits, hypothesis splits, or any run where multiple workers can move without blocking each other.

Several branches must all finish before the next step can continue.

Use for all-tests-pass checkpoints, package release trains, or situations where the next action needs the complete set of outputs.

Several branches arrive, and the output is a synthesized conclusion rather than a simple continuation.

Use for research, diagnosis, source mining, review consolidation, and comparing several proposed patches before choosing what to keep.

Several possible branches exist, but one path is selected.

Use for performance experiments, design variants, algorithm choices, and "try several approaches, keep the best one" workflows.

Continuation is blocked until a policy or check passes.

Use for automated tests, human decisions, budget checks, evidence thresholds, merge admission, and release approval. Retry behavior belongs in a loop around the gate.

The system runs a step, checks it at a gate, continues on pass, and returns to the step on fail.

Use for debugging, rerunning with tighter context, benchmark tuning, and repair-until-passing workflows.

Workflows

One worker attempts the task, then a gate decides whether the result is good enough. Failure returns to the same worker shape with better evidence or a narrower instruction.

Source fans out to V1, V2, and V3. Select chooses the strongest result, then a gate decides whether it is ready. Failure returns to the source so the next round can try a different split or better constraints.

Source fans out to V1, V2, and V3. Reduce turns the separate outputs into one conclusion before the gate judges it.

Source expands through a tree of smaller questions. The leaves reduce into a single answer, and the gate checks whether the synthesis is strong enough to use.