tama 玉

Free, open-source, code-free orchestration for multi-agent workflows.

Quickstart · Hello World · Examples

If you want to quickly try an agent flow, why are you still setting up a Python project, wiring graph nodes, defining state types, and writing routing functions before you can even iterate on the prompts?

That is the problem tama is trying to solve.

With tama, agents and skills are just Markdown files:

• agents live in AGENT.md
• skills live in SKILL.md
• orchestration is declared in YAML frontmatter
• routing can be an explicit FSM instead of hidden in code

By "code-free," we mean no graph/orchestration code for the workflow itself. You define the system in files instead of assembling it in Python.

Why tama

• No scaffold code for agent flows. Use tama init and tama add, then start writing prompts.
• Prompts as files. Agents and skills are human-readable, diffable, and easy to reorganize.
• Deterministic routing. Use fsm when control flow should belong to the runtime, not the model.
• Built-in patterns. react, fsm, scatter, critic, reflexion, debate, plan-execute, and more.
• Tracing included. Inspect which agents ran, which tools were called, and which skills were loaded.
• Rust runtime. tamad is a native binary, not a Python orchestrator.

What it looks like

This is a real workflow:

---
name: support
pattern: fsm
initial: triage
states:
  triage:
    - billing: billing-agent
    - technical: tech-agent
  billing-agent:
    - done: ~
    - escalate: triage
  tech-agent: ~
---

Instead of writing routing code, you declare the transitions.

Quick start

tama init my-project
cd my-project
tama add fsm support
tama add react triage
tama add react billing-agent
tama add react tech-agent

Then edit the generated AGENT.md files and run:

tama run "Customer says they were double charged and want a refund"

Or start with the docs:

• Quickstart: https://tama.mlops.ninja/getting-started/quickstart/
• Hello World: Deep Research: https://tama.mlops.ninja/getting-started/hello-world-deep-research/

A more complete example

The best current example is the deep research workflow in examples/00-deep-research.

It combines:

• fsm for the outer review loop
• scatter for fan-out research
• react workers for focused web research
• memory for retry-aware state
• files for writing report.md

Read the step-by-step walkthrough here:

https://tama.mlops.ninja/getting-started/hello-world-deep-research/

Current CLI

tama init <name>              # create a new project
tama add <pattern> <name>     # scaffold an agent
tama add skill <name>         # scaffold a skill
tama lint                     # validate the project
tama run "your task"          # execute the entrypoint agent

Example project structure

my-project/
├── tama.toml
├── agents/
│   └── my-project-agent/
│       └── AGENT.md
└── skills/

Larger projects usually look like:

my-project/
├── tama.toml
├── agents/
│   ├── pipeline/
│   │   └── AGENT.md
│   ├── worker/
│   │   └── AGENT.md
│   └── reviewer/
│       └── AGENT.md
└── skills/
    ├── search-web/
    │   └── SKILL.md
    └── memory/
        └── SKILL.md

Patterns

Built-in patterns currently include:

• oneshot
• react
• scatter
• parallel
• fsm
• critic
• reflexion
• constitutional
• chain-of-verification
• plan-execute
• debate
• best-of-n
• human

Docs

• Introduction: https://tama.mlops.ninja/getting-started/introduction/
• Installation: https://tama.mlops.ninja/getting-started/installation/
• Quickstart: https://tama.mlops.ninja/getting-started/quickstart/
• Hello World: Deep Research: https://tama.mlops.ninja/getting-started/hello-world-deep-research/

Status

tama is usable now, but still early.

The best way to help is to:

• try a real workflow
• report DX pain
• file issues when something is unclear or broken
• tell us which examples feel useful and which feel toy-like

License

MIT