abtreeBehaviour trees for AI agents

Define agent workflows as YAML trees. The runtime hands the agent one step at a time, verifies the result, and persists the cursor — so workflows stay reproducible no matter how big they get.

Get started

Why behaviour trees?

One step at a time

Agents see only the next instruction, not the full plan. No 2,000-line prompts. No "jumping ahead". No skipped invariants.

Durable by default

Every execution persists as a JSON document. Resume work hours or weeks later — the cursor remembers where you left off.

Visual execution traces

Every state change regenerates a Mermaid diagram. Green for success, red for failure. See exactly what your agent did and where it stopped.

Framework-agnostic

A CLI any agent can drive. Works with Claude, ChatGPT, local models — anything that can run a shell command.

from simple and ambiguousto complex and specific

simple and ambiguous

simple.yaml

seqDeploy_Service

├─ Run_Tests

├─ Build_Image

└─ Push_Image

complex and specific

deploy.yaml

seqDeploy_Service

├─ seqTest_Suite

│ ├─ Lint

│ ├─ Unit_Tests

│ └─ Integration_Tests

├─ selBuild_Strategy

│ ├─ Try_Cached_Build

│ └─ Full_Build

├─ parVerify

│ ├─ Smoke_Test

│ └─ Health_Check

└─ Push_Image

define a treethe agent drives execution

define a tree

version: 1.0.0

tree:

type: sequence

children:

-type: action

steps:

-instruct: |

Run tests.

Store pass/fail at $LOCAL.tests_passed.

Store coverage percentage at $LOCAL.coverage.

-evaluate: |

$LOCAL.tests_passed is true.

$LOCAL.coverage is greater than $GLOBAL.threshold.

-type: action

steps:

-instruct: |

Build and push image to $GLOBAL.registry.

Store the pushed tag at $LOCAL.image_tag.

state:

local:

tests_passed: null

coverage: null

image_tag: null

global:

threshold: 80

registry: ghcr.io/my-app

the agent drives execution

What is abtree?

Modern LLMs follow Markdown instructions remarkably well — until workflows grow. Then two things go wrong:

Instruction fatigue. A long system prompt loses focus. Agents skim, skip steps, hallucinate fields.
Non-determinism. Decisions left to the model produce different paths on every run.

abtree is a CLI that solves both. You describe the workflow once as a YAML behaviour tree — the same formal structure game AI and robotics have used for decades. Your agent then drives the execution through three commands:

abtree next — ask the runtime what to do.
abtree eval — answer a precondition (true or false).
abtree submit — report the outcome of an instruction.

Every call returns JSON, advances the cursor, and persists the execution to disk. The agent only ever sees the next step. The path stays predictable. The state survives.

Get started in five minutes →