Engine SDK¶

The engine is the autonomous execution loop built into antcrew. It drives an EngineLoop over a set of capabilities until a goal is satisfied — no fixed pipeline, no manual step sequencing.

pip install antcrew   # antcrew_engine is bundled — no separate install needed

Note: antcrew-engine was a separate package until antcrew v0.35.0. It is now merged into antcrew. All imports (from antcrew_engine import EngineLoop) continue to work unchanged.

What it provides¶

EngineLoop — the decision loop. On each iteration it inspects the current project state, selects the best-fit capability (Architect, CodeGenerator, TestRunner…), dispatches it, and checks whether the goal conditions are now satisfied.

Built-in capabilities — Architect, TaskPlanner, CodeGenerator, TestGenerator, TestRunner, BugFixer, CodeReviewer, DocGenerator, SecurityScanner, and more. Each capability reads from the ArtifactStore and writes typed, Pydantic-validated artifacts back to it.

Documentation Module — index your project's existing docs (markdown, Word, PDF, Jira, COBOL) and inject relevant context into every capability's LLM prompt via --schema / --docs-dir on the CLI, or DocumentationManager in the Python API.

Legacy / COBOL Support — first-class support for IBM AS/400 organisations: parse .cbl/.cpy files, connect to DB2 for i via AS400Connector, add AI to COBOL programs with COBOLAugment, or translate programs to Python/Java/Go with polytranslate.

Provider-agnostic model calls — change "claude:claude-sonnet-5" to "openai:gpt-4o" and nothing else changes. build_llm() resolves the string to the correct provider client.

EventLog — every capability dispatch, result, and retry is written to a structured, append-only event log. antcrew-platform receives these events in real time and shows them in the Runs dashboard.

HITL checkpoints — HitlReviewer is a built-in capability that pauses the loop and sends a review request to antcrew-platform. Execution resumes once a human approves or rejects from the dashboard.

Quick start¶

from antcrew_engine import (
    EngineLoop, MemoryStore, Goal, DesiredProjectState,
    Constraints, Condition, ConditionId,
    Architect, TaskPlanner, CodeGenerator, TestGenerator, TestRunner,
    artifact_validators,
)
from antcrew_engine import CapabilityRegistry
from antcrew_engine.config import build_llm
from antcrew_engine.engine import EventLog

# 1. Configure the LLM
llm = build_llm("claude:claude-sonnet-5")

# 2. Register capabilities
registry = CapabilityRegistry()
registry.register(Architect(llm))
registry.register(TaskPlanner(llm))
registry.register(CodeGenerator(llm))
registry.register(TestGenerator(llm))
registry.register(TestRunner())

# 3. Define the goal
goal = Goal(
    description="Build a JWT authentication module",
    desired_state=DesiredProjectState(
        conditions=[
            Condition(ConditionId("architecture_exists"), "Architecture document produced"),
            Condition(ConditionId("implementation_exists"), "All tasks have code files"),
            Condition(ConditionId("tests_pass"), "Test suite passes"),
        ]
    ),
    constraints=Constraints(max_iterations=20),
)

# 4. Run
store = MemoryStore()
event_log = EventLog()
engine = EngineLoop(registry, artifact_validators, event_log)
final_state = engine.run(store, goal)

Connecting to the platform¶

To stream events to antcrew-platform in real time, pass a PlatformEventBridge as the event log:

from antcrew_engine.engine import EventBusBridge

bridge = EventBusBridge(
    platform_url="https://antcrew.org",
    api_key="acw_live_...",
    run_id="your-run-id",
)
engine = EngineLoop(registry, artifact_validators, bridge)

Every capability dispatch and result will appear in the platform dashboard under the run's event timeline.

Core concepts¶

Concept	What it is
`EngineLoop`	The decision loop — selects and dispatches capabilities until the goal is satisfied
`Capability`	A discrete unit of work (Architect, CodeGenerator, TestRunner…) that reads and writes typed artifacts
`ArtifactStore`	In-memory (`MemoryStore`) or filesystem (`FilesystemStore`) store for typed artifacts
`Goal`	The target state the engine works toward, expressed as `Condition` objects
`EventLog` / `EventBusBridge`	Structured log of every engine event; bridged to the platform for live observability
`build_llm(model)`	Factory that resolves a model string to a provider-specific LLM client

Typed artifacts

EventLog & Replay

LLM providers