acty

Async grouped-job runtime for Python, AI agents, and LLM systems, with retries, fair scheduling, JSONL event logs, and a live terminal TUI.

Acty is a good fit when one setup step unlocks a family of related async jobs. That can be an AI agent that establishes session state, a coordinator that fans work out to specialist agents, an LLM request with shared context, a cache warmup step for prefix caching, or any workload where several follow-up jobs should stay tied to the same unit of work.

It builds on acty-core and adds the high-level engine API, the TUI app, and the acty-tui CLI for following or replaying JSONL event streams.

Table of Contents

• Why use acty?
• Core ideas
• Why primer/follower instead of a flat task queue?
• Install
• Quick Start
• LangChain With acty-langchain
• Open Groups + Cache Warmup
• Lanes
• Watch A Run Live
• Examples
• Package Map
• Development

Why use acty?

• Keep related jobs together instead of manually stitching queues, futures, and lifecycle state.
• Run one primer job and then fan out follower jobs behind it.
• Retry work without bolting retry logic onto every executor.
• Split worker capacity fairly across runs, tenants, or pipelines with lanes.
• Write JSONL events and inspect long runs live in a terminal TUI.

Core ideas

• group: one logical unit of work
• primer: the setup step for a group, usually the job that establishes context, warms a cache, creates or reuses a provider/session handle, or prepares shared inputs
• follower: work that belongs to the same group and usually benefits from that setup step
• lane: a fairness label that gives a run or tenant its own share of worker capacity

If you have ten noisy runs and one important run, lanes let you give the important run a bigger share instead of letting everything fight in one flat queue. Idle lanes do not have to waste capacity; other lanes can still use spare workers.

Why primer/follower instead of a flat task queue?

Because a lot of real systems are not just "N independent tasks". They usually have a shape more like:

1. Resolve or create shared context.
2. Fan out follow-up work that depends on that context.

That first step might be:

• opening or reusing a provider session
• warming a prompt or session cache for prefix caching before fan-out
• loading shared memory or tool context for an AI agent
• building a cache handle or provider reference
• doing one expensive planning step before parallel execution

In a flat queue, that structure is implicit. You end up rebuilding it yourself with ad-hoc dependency tracking, custom cancellation rules, and queueing hacks. Acty makes that boundary explicit so the runtime can do useful things with it:

• schedule setup work separately from fan-out work
• avoid a large follower burst starving new groups from getting their first step started
• keep retries and failure policies aligned with the actual shape of the workflow
• attach cache/session state to the group instead of smearing it across unrelated tasks
• model cache warmup explicitly when the first call prepares prefix-caching state for later calls
• make open groups natural when more followers arrive later

The point is not that every workload must have a dramatic "leader" task. The point is that many agent and LLM systems have a setup phase and an execution phase, and Acty gives those phases names the scheduler can understand.

Install

pip install acty

Optional LangChain integration:

pip install "acty[langchain]"

Local development:

pip install -e .[dev]

Quick Start

This example shows a more realistic grouped workflow: one coordinator step sets up the job, then several specialist steps continue the same unit of work.

import asyncio
from pprint import pprint

from acty import ActyEngine, EngineConfig, EchoExecutor


async def main() -> None:
    # EchoExecutor is a zero-dependency demo executor: it simply returns the
    # payload it received. That makes it useful for learning the Acty API
    # without bringing in a model provider yet.
    engine = ActyEngine(executor=EchoExecutor(), config=EngineConfig())
    try:
        # The primer carries the shared setup and the full initial context.
        submission = await engine.submit_group(
            "release-readiness",
            {
                "agent": "coordinator",
                "task": "set up the shared review context",
                "messages": [
                    {
                        "role": "system",
                        "content": "You coordinate release-readiness reviews for an engineering team.",
                    },
                    {
                        "role": "user",
                        "content": "We want to ship the billing service update today.",
                    },
                    {
                        "role": "user",
                        "content": "Figure out which risks and checks should be reviewed before rollout.",
                    },
                ],
            },
            [
                {
                    "agent": "specialist:test-risk",
                    "task": "review regression risk",
                    "messages": [
                        {
                            "role": "user",
                            "content": "Look for missing tests or weak coverage before release.",
                        },
                    ],
                },
                {
                    "agent": "specialist:performance",
                    "task": "review latency and throughput risk",
                    "messages": [
                        {
                            "role": "user",
                            "content": "Check whether the release could hurt latency or throughput.",
                        },
                    ],
                },
                {
                    "agent": "specialist:ops",
                    "task": "review rollout and monitoring readiness",
                    "messages": [
                        {
                            "role": "user",
                            "content": "Check whether the deployment plan and monitoring are good enough.",
                        },
                    ],
                },
            ],
        )

        # Followers only carry the incremental asks that branch off that setup.
        if submission.primer is not None:
            pprint({"primer": (await submission.primer).output})
        for index, future in enumerate(submission.followers, start=1):
            pprint({"follower": index, "output": (await future).output})
    finally:
        await engine.close()


asyncio.run(main())

Notice that the followers only add the next request. They do not need to repeat the whole shared setup. With a real executor, they would typically reuse the group context, provider session, or cache state prepared by the primer. EchoExecutor just makes the grouped payloads easy to inspect.

LangChain With acty-langchain

Install the integration and the core LangChain package:

pip install acty acty-langchain langchain-core

If you want to plug in a real chat model, also install the matching provider integration, for example:

pip install langchain-openai

This example stays provider-free by using RunnableLambda, but the chain shape is the same one you would use with a real LangChain chat model.

import asyncio
from pprint import pprint

from acty import ActyEngine, EngineConfig
from acty_langchain import LangChainRunnableExecutor
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.runnables import RunnableLambda


def fake_chat_model(prompt_value):
    # RunnableLambda stands in for a real chat model so the example stays
    # runnable without API keys. In production, replace this with ChatOpenAI,
    # another chat model, or a longer LangChain pipeline.
    messages = prompt_value.to_messages()
    return {
        "message_count": len(messages),
        "last_message": messages[-1].content if messages else "",
    }


async def main() -> None:
    # This is a normal LangChain chat chain: prompt template first, then the
    # model step. Acty does not replace LangChain here; it runs the chain as the
    # job executor payload.
    review_chain = (
        ChatPromptTemplate.from_messages(
            [
                ("system", "You help coordinate release reviews for engineering teams."),
                ("human", "Shared context: {shared_context}"),
                ("human", "Current task: {task}"),
            ]
        )
        | RunnableLambda(fake_chat_model)
    )

    engine = ActyEngine(
        # LangChainRunnableExecutor reads `runnable` and `input` from each Acty
        # payload and invokes the LangChain chain for that job.
        executor=LangChainRunnableExecutor(),
        config=EngineConfig(),
    )
    try:
        # The primer handles the broad coordination step for the group.
        # Followers reuse the same chain but branch into narrower review tasks.
        submission = await engine.submit_group(
            "langchain-release-review",
            {
                "runnable": review_chain,
                "input": {
                    "shared_context": "Billing service release planned for today.",
                    "task": "Plan the review and identify the main risk areas.",
                },
            },
            [
                {
                    "runnable": review_chain,
                    "input": {
                        "shared_context": "Billing service release planned for today.",
                        "task": "Check automated-test risk for this release.",
                    },
                },
                {
                    "runnable": review_chain,
                    "input": {
                        "shared_context": "Billing service release planned for today.",
                        "task": "Check performance risk for this release.",
                    },
                },
                {
                    "runnable": review_chain,
                    "input": {
                        "shared_context": "Billing service release planned for today.",
                        "task": "Check rollout and alerting readiness.",
                    },
                },
            ],
        )

        if submission.primer is not None:
            pprint({"primer": (await submission.primer).output})
        for index, future in enumerate(submission.followers, start=1):
            pprint({"follower": index, "output": (await future).output})
    finally:
        await engine.close()


asyncio.run(main())

Use this pattern when you already have a LangChain chat chain and want Acty to manage grouped execution, retries, event logs, and live monitoring around it.

Open Groups + Cache Warmup

Use open_group(...) when the work for a group arrives over time. This is a good fit for prefix caching: warm the shared cache/session once with the primer, then keep sending followers that reuse the same handle.

import asyncio
from pprint import pprint

from acty import ActyEngine, EngineConfig
from acty_core.cache import CacheRegistry, CacheProvider, InMemoryStorage
from acty_core.core.types import Job, JobResult


class DemoPrefixCacheProvider:
    # This fake provider models what a real prefix-cache/session provider would
    # do for Acty: derive a stable cache identity from the primer payload and
    # create a provider_ref that later jobs can reuse.
    name = "demo-prefix"

    def fingerprint(self, primer, context=None) -> str:
        tenant = context["tenant"] if context is not None else "default"
        prefix_id = primer["prefix_id"]
        return f"{tenant}:{prefix_id}"

    async def create(self, primer, context=None) -> str | None:
        tenant = context["tenant"] if context is not None else "default"
        prefix_id = primer["prefix_id"]
        return f"prefix-cache:{tenant}:{prefix_id}"

    async def wait_ready(self, provider_ref, timeout_s=None) -> None:
        return None

    async def invalidate(self, provider_ref) -> None:
        return None


class PrefixAwareExecutor:
    # This fake executor does not call a model. It only exposes the shared
    # provider_ref in the result so you can see that the same cache/session
    # handle reaches every job in the group.
    async def execute(self, job: Job, *, pool: str) -> JobResult:
        cache = job.context.get("cache") if job.context is not None else None
        provider_ref = cache.get("provider_ref") if cache is not None else None
        return JobResult(
            job_id=job.id,
            kind=job.kind,
            ok=True,
            output={
                "agent": job.payload["agent"],
                "provider_ref": provider_ref,
                "last_message": job.payload["messages"][-1]["content"],
            },
            group_id=job.group_id,
            follower_id=job.follower_id,
        )


async def main() -> None:
    # CacheRegistry is the piece that remembers cache/session handles across the
    # group. InMemoryStorage keeps the example self-contained.
    registry = CacheRegistry(DemoPrefixCacheProvider(), storage=InMemoryStorage())
    engine = ActyEngine(
        executor=PrefixAwareExecutor(),
        config=EngineConfig(),
        cache_registry=registry,
    )
    try:
        # open_group(...) is useful here because the whole job does not arrive
        # up front: we warm the cache first, then send more work into the same
        # group as new follow-up questions appear.
        async with engine.open_group(
            "support-playbook-review",
            # cache_key says "all jobs in this group belong to the same warmed
            # prefix/session entry".
            cache_key="tenant:acme:support-playbook-v3",
            # cache_context carries any extra fields the provider needs to build
            # or look up that cache/session entry.
            cache_context={"tenant": "acme"},
        ) as group:
            # The primer warms the shared prefix/session state once for the group.
            primer = await group.submit_primer(
                {
                    "agent": "coordinator",
                    "prefix_id": "support-playbook-v3",
                    "messages": [
                        {
                            "role": "system",
                            "content": "Shared support playbook, escalation policy, and customer-history rules.",
                        },
                        {
                            "role": "user",
                            "content": "Warm the shared context for this investigation.",
                        },
                    ],
                }
            )

            # Followers can arrive later and still reuse the same warmed handle.
            followers = [
                await group.add_follower(
                    {
                        "agent": "specialist:billing",
                        "prefix_id": "support-playbook-v3",
                        "messages": [
                            {
                                "role": "user",
                                "content": "Check invoice retry failures for this customer.",
                            },
                        ],
                    }
                ),
                await group.add_follower(
                    {
                        "agent": "specialist:webhooks",
                        "prefix_id": "support-playbook-v3",
                        "messages": [
                            {
                                "role": "user",
                                "content": "Check whether webhook delivery is delayed or failing.",
                            },
                        ],
                    }
                ),
                await group.add_follower(
                    {
                        "agent": "specialist:history",
                        "prefix_id": "support-playbook-v3",
                        "messages": [
                            {
                                "role": "user",
                                "content": "Pull recent escalations that look similar.",
                            },
                        ],
                    }
                ),
            ]

        if primer is not None:
            pprint({"primer": (await primer).output})
        for index, future in enumerate(followers, start=1):
            pprint({"follower": index, "output": (await future).output})
    finally:
        await engine.close()


asyncio.run(main())

All results will show the same provider_ref. That is the point of the group here: the runtime has one place to attach the warmed cache/session handle and propagate it to every job in the same unit of work.

With a real cache-aware executor, that shared handle can mean one of two things:

• the primer warms a provider-side prefix cache and followers hit that same warmed prefix
• the primer creates or reuses a session/provider reference and followers continue from it without starting cold

That is much harder to model cleanly with a flat queue of unrelated tasks.

Lanes

Lanes are how you give different workloads a fair share of the same worker pool.

Typical uses:

• separate traffic by tenant
• protect one evaluation run from another
• keep background batch work from swallowing all follower workers

You can attach lanes either by setting group_context={"lane": ...} when you submit a group or by using engine.lane(...) helpers.

The important part is the relative weight:

• weight=2 means "this lane should get about twice the share"
• weight=1 means "normal share"

That gives you fairness without hard-partitioning the workers into isolated pools.

Watch A Run Live

Acty works well when you want both a programmatic API and a live view of what the runtime is doing. The example below creates a richer event stream: multiple groups, multiple followers, and two lanes competing in the same run.

Write JSONL events while several groups run across two lanes:

import asyncio
from pathlib import Path

from acty import ActyEngine, EngineConfig, SimulatedExecutor
from acty_core.events.sinks.jsonl import JsonlEventSink
from acty_core.scheduler import LaneConfig


def build_followers(*names: str) -> list[dict]:
    # Keep follower payloads tiny here: the goal of this example is to generate
    # visible scheduling activity, not to show prompt design. Each returned
    # dict becomes one follower job inside a group.
    return [
        {
            "agent": name,
            "messages": [{"role": "user", "content": name}],
        }
        for name in names
    ]


async def main() -> None:
    # JsonlEventSink writes runtime events to disk so the TUI can follow them
    # live or replay them later.
    sink = JsonlEventSink(Path("/tmp/acty_events.jsonl"))
    engine = ActyEngine(
        # Use simulated delays so the TUI has enough activity to watch.
        executor=SimulatedExecutor(min_delay_s=0.2, max_delay_s=0.8, seed=7),
        config=EngineConfig(
            primer_workers=2,
            follower_workers=3,
            # Give the priority lane about twice the scheduling share of the
            # background lane when both have queued work.
            lane_configs={
                "priority": LaneConfig(weight=2),
                "background": LaneConfig(weight=1),
            },
        ),
        event_sinks=[sink],
    )
    try:
        # Submit several groups at once so the event log shows queueing, execution,
        # and lane competition instead of a single short burst.
        submissions = await asyncio.gather(
            engine.submit_group(
                "priority:billing",
                # Keep messages short so the example stays readable; the point is
                # to create several groups that compete for runtime capacity.
                {"agent": "coord", "messages": [{"role": "user", "content": "plan"}]},
                build_followers("tests", "perf", "ops"),
                group_context={"lane": "priority"},
            ),
            engine.submit_group(
                "priority:search",
                {"agent": "coord", "messages": [{"role": "user", "content": "plan"}]},
                build_followers("tests", "perf", "alerts"),
                group_context={"lane": "priority"},
            ),
            engine.submit_group(
                "background:analytics",
                {"agent": "coord", "messages": [{"role": "user", "content": "plan"}]},
                build_followers("schema", "perf", "dash"),
                group_context={"lane": "background"},
            ),
            engine.submit_group(
                "background:ops",
                {"agent": "coord", "messages": [{"role": "user", "content": "plan"}]},
                build_followers("logs", "alerts", "runbook"),
                group_context={"lane": "background"},
            ),
        )

        # Wait for every primer and follower so the JSONL file captures the whole
        # run instead of stopping after only a few early events.
        pending = []
        for submission in submissions:
            if submission.primer is not None:
                pending.append(submission.primer)
            pending.extend(submission.followers)
        await asyncio.gather(*pending)
    finally:
        await engine.close()
        sink.close()


asyncio.run(main())

This produces a much better event stream for the TUI: multiple groups, queued and completed followers, and lane-tagged work competing in the same run.

Then watch the event stream:

acty-tui follow /tmp/acty_events.jsonl

Or replay a finished run:

acty-tui replay /tmp/acty_events.jsonl --speed 2.0

Examples

• python examples/group_tui_demo.py --groups 3 --followers-per-group 2
• python examples/lane_demo.py
• python examples/tenacity_retry_demo.py --event-jsonl /tmp/acty_tenacity_events.jsonl

Package Map

• acty-core: low-level scheduler, lifecycle, event, and cache primitives
• acty: engine, client, TUI, and CLI
• acty-langchain: LangChain payload helpers and runnable executor support
• acty-openai: OpenAI executor integration
• acty-gigachat: GigaChat executor integration

If you only need the lower-level runtime pieces, use acty-core. If you want the higher-level API and the live terminal tooling, use acty.

Development

• tests live under tests/
• example programs live under examples/
• adapter-specific integration coverage belongs in the adapter packages