Patterns Overview

tama has 13 built-in patterns. Every pattern is a pre-wired composition of LLM calls; you declare the pattern and write the prompts — tama handles the control flow.

Quick reference

Pattern	Complexity	Required files	Best for
oneshot	⬤	(none)	Simple transformations, classification, formatting
react	⬤⬤	(none)	Tool use, research, multi-step reasoning
scatter	⬤⬤⬤	reduce.md	Parallel research, batch processing
parallel	⬤⬤	(none)	Independent analyses, voting
fsm	⬤⬤⬤	(none)	Custom workflows, conditional routing
critic	⬤⬤	draft.md, critique.md, refine.md	Writing quality, iterative improvement
reflexion	⬤⬤⬤	reflect.md	Self-correcting agents, performance improvement
constitutional	⬤⬤	critique.md, revise.md	Safe generation, principle-adherent output
chain-of-verification	⬤⬤⬤	verify.md, check.md, revise.md	Factual accuracy, claim verification
plan-execute	⬤⬤⬤	execute.md, verify.md	Complex tasks with verifiable sub-steps
debate	⬤⬤⬤	(none — uses named agents)	Balanced analysis, decision-making
best-of-n	⬤⬤⬤	judge.md	High-quality generation through selection
human	⬤⬤⬤	resume.md	Human-in-the-loop approval, interactive tasks

How patterns are built

Under the hood, all patterns are combinations of two primitives:

FSM      — sequential steps with routing
Parallel — concurrent steps with collection

The named patterns are pre-wired compositions so you don’t have to write the FSM YAML yourself:

critic     = FSM: draft → critique → refine
scatter    = FSM(map) → Parallel(workers) → FSM(reduce)
best-of-n  = Parallel(N variants) → FSM(judge)
debate     = FSM: position-a → position-b → judge

Use fsm directly when none of the named patterns fit your workflow.

Choosing a pattern

Start simple. Most tasks fit oneshot or react:

• Need just an LLM response? → oneshot
• Need to use tools or reason in multiple steps? → react
• Need to improve quality iteratively? → critic or reflexion
• Need to process many items in parallel? → scatter
• Need factual accuracy? → chain-of-verification
• Need a complex custom workflow? → fsm

Step files

Patterns with multiple steps load system prompts from separate .md files. These files live alongside AGENT.md in the agent directory and can optionally include their own call: frontmatter to override the model or tools for that specific step:

agents/
  essay-critic/
    AGENT.md        ← pattern: critic, shared config
    draft.md        ← system prompt for the draft step
    critique.md     ← system prompt for the critique step
    refine.md       ← system prompt for the refine step

Each step file can optionally include frontmatter:

---
call:
  model:
    name: anthropic:claude-opus-4-6
    temperature: 0.2
---

You are a meticulous critic. Identify the three most significant weaknesses...

See Step files reference for details.