Cross Agents Runtime (Python)

Cross Agents Runtime is a provider-agnostic, model-adaptive agent runtime for building controlled mono-agent and multi-agent systems on Python. It picks an execution pattern that fits a given task, model, policy, and operational constraints, then runs it inside an audited, bounded session.

This repository contains the Python implementation of the framework. Implementations in other ecosystems (.NET, TypeScript) live in separate repositories so each can follow the conventions and release cadence of its own ecosystem.

Installation

The package is published to PyPI at version 0.1.0a1 (alpha pre-release). Requires Python >= 3.11.

# pip - alpha versions need --pre
pip install --pre crossagents-runtime

# or pin the version explicitly
pip install crossagents-runtime==0.1.0a1

# uv - alpha versions need --prerelease=allow
uv add crossagents-runtime --prerelease=allow

After install:

from crossagents.core import AgentRuntime, RuntimeOptions
from crossagents.patterns import NoToolPattern, PlanExecuteValidatePattern
from crossagents.testing import FakeModelAdapter, InMemoryAuditSink

What it is

• A small set of stable contracts (crossagents.abstractions) describing models, tools, memory, patterns, policy, and audit.
• A minimal runtime host (crossagents.core) that registers adapters and patterns, selects one for each task, runs it, and surfaces a structured result.
• A first-party set of safe patterns (crossagents.patterns): a no-tool single call, a Plan-Execute-Validate flow, a JSON plan skeleton, and a strictly bounded ReAct loop.
• Optional layers for tooling (crossagents.tooling) and memory (crossagents.memory) that can be plugged in independently.
• Deterministic test doubles (crossagents.testing) for pattern and runtime tests with no external dependencies.

What it isn't

• Not a chatbot framework. Conversations are a use case applications can build on top, not the framework's purpose.
• Not coupled to a single LLM provider. The framework ships no provider adapters in this milestone; applications wire their own through IModelAdapter.
• Not a tool-calling library. Tooling is an optional module; tasks that don't need tools never see the tooling layer.
• Not a memory store. Memory retrieval is an optional module; the framework does not own a database or vector index.
• Not unbounded. Patterns must declare bounds. Unbounded ReAct configurations are rejected at registration time.

Core concepts

• Model adapter: a thin IModelAdapter implementation that talks to one model and reports ModelCapabilities.
• Agent task: a single unit of work (AgentTask) with a type, input, and optional requirements.
• Pattern: an IAgentPattern plus a PatternDescriptor that declares its risk profile, bounds, and dependencies.
• Pattern selector: a deterministic chooser that filters patterns by task requirements, model capabilities, and policy, then scores survivors.
• Policy engine: an IPolicyEngine that translates declarative policies (AgentPolicy) into yes/no decisions for selection and tool calls.
• Audit pipeline: a per-session buffer that fans events out to IAuditSink and surfaces them in the runtime result.

Minimal example

import asyncio

from crossagents.abstractions.agents import AgentTask
from crossagents.abstractions.models import (
    ModelCapabilities, ModelProfile, ModelProvider,
)
from crossagents.core import AgentRuntime, RuntimeOptions
from crossagents.patterns import NoToolPattern, PlanExecuteValidatePattern
from crossagents.testing import FakeModelAdapter, InMemoryAuditSink


async def main() -> None:
    runtime = AgentRuntime(RuntimeOptions(audit_sink=InMemoryAuditSink()))

    profile = ModelProfile(
        profile_id="demo-echo",
        display_name="Demo echo model",
        provider=ModelProvider.CUSTOM,
        capabilities=ModelCapabilities(
            provider_name="demo", model_id="echo", is_local=True,
        ),
    )

    runtime.register_model(
        FakeModelAdapter(profile, "Hello from Cross Agents Runtime.")
    ).register_pattern(NoToolPattern()).register_pattern(PlanExecuteValidatePattern())

    result = await runtime.run(
        AgentTask(task_id="demo-1", input="Say hello.", requires_validation=False),
        profile.profile_id,
    )

    print(f"{result.selected_pattern_id}: {result.agent.output if result.agent else ''}")


if __name__ == "__main__":
    asyncio.run(main())

A self-contained runnable version of this lives in examples/minimal_runtime.py.

Package layout

Module	Purpose
crossagents.abstractions	Stable contracts (models, tools, memory, patterns, policy, audit)
crossagents.core	Runtime host, session, selector, default policy engine, audit pipeline
crossagents.patterns	First-party safe patterns
crossagents.tooling	Optional tool registry, validator, executor, normalizer
crossagents.memory	Optional retrieval, ranking, compression, sliding buffer
crossagents.testing	Deterministic test doubles

Design principles

1. Provider-agnostic: model behaviour reaches the runtime only through IModelAdapter and ModelCapabilities.
2. Model-adaptive: pattern selection inspects the model's capability profile and degrades safely when something is missing.
3. Bounded by default: every shipped pattern declares step counts and risk levels; the runtime rejects unbounded configurations.
4. Optional middleware: tooling and memory are separate modules and separate runtime services; they can be omitted entirely.
5. Auditable: every session emits a canonical sequence of audit events suitable for compliance and debugging.
6. Deterministic to test: crossagents.testing ships in-process fakes for every external dependency the framework defines.
7. Small public surface: contracts are short, immutable, and documented; framework code never exposes provider-specific types.

Current status

This is the first milestone. It establishes the contracts, the runtime, three patterns plus a strictly bounded ReAct loop, optional tooling and memory layers, and the test surface. Provider adapters, multi-agent orchestration primitives, distributed session storage, and other extensions are out of scope for this milestone.

What is intentionally not in this milestone

• No real provider adapters (OpenAI, Anthropic, Bedrock, Ollama, etc.).
• No vector store, embedding, or document extraction implementations.
• No multi-agent orchestration patterns (debate, swarm, voting). The contracts allow them; an implementation will arrive in a later milestone.
• No streaming response surface beyond the boolean capability flag.
• No long-term session persistence.
• No telemetry exporter (open-telemetry, Prometheus, etc.). The audit sink is the integration point.

Building and testing

Using uv (recommended):

uv sync --all-extras
uv run pytest
uv run python examples/minimal_runtime.py
uv build

Using pip and python -m build:

python -m pip install -e ".[test]"
pytest
python examples/minimal_runtime.py
python -m pip install build && python -m build

Tests run entirely in-process and require no credentials or network access.

License

MIT. See LICENSE.