Agently 4.1.3.7 - AI Application Runtime Framework

Build AI service backends with structured outputs, observable Actions, runtime Skills, MCP capabilities, process streams, and recoverable workflows.

English | 中文介绍

Docs · Quickstart · Why Agently · Capabilities · Architecture · Ecosystem

Who This README Is For

Agently is for teams moving from "the model can do it once" to "the application must do it reliably":

• product engineers building assistants, internal copilots, knowledge tools, operation workflows, or AI-backed APIs
• platform teams that need clear extension points for model providers, tools, MCP servers, sandboxes, workflows, and observability
• technical leads comparing AI frameworks for maintainability, explicit control, debuggability, and production handoff
• coding-agent users who want a framework whose recommended patterns can be encoded as reusable project guidance

The main design question is simple: how do you keep model behavior useful while still giving application code stable contracts, observable execution, and restart-safe workflow boundaries?

Agently 4.1.3.7 hardens the AgentExecution-backed AgentTaskLoop path: goal and success criteria stay on one execution draft, effort(...) controls bounded strategy depth, DAG-shaped steps feed evidence back into verification, Skills context packs can support planners and DAG consumers, and ActionRuntime release blockers are closed without changing ownership boundaries. Read the 4.1.3.7 Release Notes, 4.1.3.6 Release Notes, 4.1.3.5 Release Notes, 4.1.3.4 Release Notes, 4.1.3.3 Release Notes, 4.1.3.2 Release Notes, 4.1.3.1 Release Notes, and 4.1.3 Release Notes for the full release story.

Why Agently

Many AI frameworks are strong at exploration or at assembling broad integration stacks. Agently is optimized for the engineering layer that makes model applications survive model changes, output drift, streaming UX, action execution, workflow signals, and service boundaries.

Agently is a good fit when you care about:

• AI services should be runtime executions, not prompt glue - one Agent execution can declare candidate Actions, Skills, MCP services, Dynamic Task planning, process streams, and output contracts, then execute through the same runtime surface. Read 4.1.3.7 Release Notes, Agent Auto Orchestration examples, and Skills Executor examples.
• Model switching should not rewrite business logic - Agently normalizes provider setup, prompt slots, response parsing, action execution, and response reading into one request/runtime contract. Read Model Setup, Models Overview, and Requests Overview.
• Structured output should be a framework guarantee, not only a provider feature - .output(...) schemas, required field extraction, parser feedback, retries, ensure_keys, ensure_all_keys, and validation handlers work together inside Agently. Read Schema as Prompt, Output Control, and examples in examples/basic/.
• Streaming should expose structure before the final token - instant mode lets consumers react to structured fields while the model is still streaming, which is useful for UI updates, SSE routes, and workflow signals. Read Model Response, FastAPI Service Exposure, and examples/fastapi/.
• Actions should be observable and model-portable - local functions, built-in actions, MCP servers, shell/Python/Node/SQLite/workspace helpers, and custom executors produce structured records and can share one Action Runtime. Read Action Runtime, MCP, and examples/action_runtime/.
• Skills should be runtime capabilities, not inline prompt snippets - agent.use_skills(...) can declare local or remote Skill sources; the Skills Executor discovers, installs, selects, mounts MCP/script capabilities, streams diagnostics, and executes only when the planner needs them. Read Skills Executor and examples/skills_executor/.
• Execution dependencies should have lifecycle owners - Execution Environment providers manage reusable resources such as MCP processes, browser sessions, shell/Python/Node runtimes, SQLite handles, and sandboxes. Read Execution Environment and examples/execution_environment/.
• Generated plans should become validated task graphs - Dynamic Task turns model-generated or app-generated DAG data into validated, observable task execution through Agently.create_dynamic_task(...). Read Dynamic Task and examples/dynamic_task/.
• Workflows should be signal-driven, not just graph-shaped - TriggerFlow supports events, fan-out, runtime streams, pause/resume, save/load, sub-flows, and close snapshots; instant structured output can become workflow input without waiting for the whole response. Read TriggerFlow Overview, Events and Streams, and examples/trigger_flow/.
• Common model-app patterns should be composable - router, To-Do/dependency execution, planning, reflection, evaluator/reviser, and multi-agent collaboration can be built from the same request/action/signal primitives. Read Playbooks, TriggerFlow Model Integration, and examples/step_by_step/.
• Services should keep clean project boundaries - async APIs, FastAPI helpers, settings files, prompt files, DevTools observation, and companion coding-agent skills fit non-trivial projects. Read Project Framework, FastAPI Service Exposure, and Observability.

Current framework version: 4.1.3.7.

Python: >=3.10.

Framework Positioning

The point is not that other frameworks are wrong. They choose different centers of gravity.

Use Agently when the application needs an AI execution substrate. Stay closer to direct SDK calls when the product is only one or two simple prompts. Use a specialized multi-agent framework when natural-language agent collaboration is the main product primitive.

The practical differences show up in four layers:

• Against LangChain's integration-first style: LangChain is strong when you want a broad, flexible set of model, tool, retrieval, and agent building blocks. Agently's bet is that production model apps need a more uniform request contract: different model providers should still feed the same prompt slots, structured parser, retry/validation path, reusable result readers, and Action Runtime. That reduces the chance that swapping the base model or provider changes the shape expected by downstream business logic. Start with Requests Overview and Action Runtime.
• Against provider-native structured output as the only guarantee: Agently can use model providers, but its output quality path does not depend only on provider-side JSON schema or tool-calling parameters. The framework owns schema-as-prompt authoring, required-field extraction, parser feedback, retries, ensure_keys, ensure_all_keys, and validation handlers. That matters when the target model does not expose the same structured-output or tool-calling semantics as another provider. See Schema as Prompt and Output Control.
• Against graph-only orchestration: LangGraph is strong for graph-shaped stateful agents and durable execution. TriggerFlow's core is event/signal-driven, and Agently's instant response mode can surface structured fields while the model is still streaming. That lets workflow signals be driven by partial structured output, action results, human input, or sub-flow state instead of waiting for a whole model response to finish. See Model Response, TriggerFlow Events and Streams, and examples/fastapi/ for streaming/service patterns.
• Against treating multi-agent as the framework root: Multi-agent collaboration is useful, but in Agently it is a scenario you can build on top of requests, Actions, TriggerFlow signals, sub-flows, Session, and runtime resources. Router, To-Do/dependency execution, planning, reflection, evaluator/reviser, and agent-team patterns are all compositions over the same lower-level engineering substrate. See Playbooks, TriggerFlow Model Integration, and examples/step_by_step/.

Quickstart

Install:

pip install -U agently

Use DeepSeek or another OpenAI-compatible hosted endpoint:

from agently import Agently

Agently.set_settings(
    "OpenAICompatible",
    {
        "base_url": "https://api.deepseek.com/v1",
        "model": "deepseek-chat",
        "auth": "DEEPSEEK_API_KEY",
        "model_type": "chat",
        "request_options": {"temperature": 0.2},
    },
)

agent = Agently.create_agent()

result = (
    agent
    .input("Introduce Python in one sentence and list three strengths.")
    .output({
        "intro": (str, "one sentence", True),
        "strengths": [(str, "one strength")],
    })
    .start(ensure_all_keys=True)
)

print(result)

Use local Ollama by changing provider settings:

ollama pull qwen2.5:7b

Agently.set_settings(
    "OpenAICompatible",
    {
        "base_url": "http://127.0.0.1:11434/v1",
        "model": "qwen2.5:7b",
        "api_key": "ollama",
        "model_type": "chat",
    },
)

Use a Model Pool when the same application needs multiple models:

agent.set_settings("model_pool", {
    "ollama-qwen2.5": "qwen2.5:7b",
    "deepseek-v4": "deepseek-chat",
})
agent.set_settings("key_pool", {
    "local": "ollama",
    "deepseek-main": "${ENV.DEEPSEEK_API_KEY}",
    "deepseek-backup": "${ENV.DEEPSEEK_BACKUP_API_KEY}",
})
agent.set_settings("key_pool_strategy", {
    "qwen2.5:7b": {"mode": "fixed", "pool": ["local"]},
    "deepseek-chat": {"mode": "round_robin", "pool": ["deepseek-main", "deepseek-backup"]},
})

agent.activate_model("ollama-qwen2.5")

activate_model(...) changes the default model key for subsequent Agent-owned requests. Use agent.create_request(model_key="deepseek-v4") for a one-off override.

For file-backed settings, prefer:

from agently import Agently

Agently.load_settings("yaml_file", "settings.yaml", auto_load_env=True)

Core Capabilities

1. Structured Requests

Prompts are composed from named slots. That keeps application intent, constraints, context, and output contracts reviewable:

result = (
    agent
    .role("You are a concise release-note writer.")
    .info({"version": "4.1.3.7", "audience": "framework users"})
    .instruct("Return only facts grounded in the input.")
    .input("Summarize this release line for an engineering changelog.")
    .output({
        "headline": (str, "short headline", True),
        "bullets": [(str, "one stable fact")],
    })
    .get_result()
)

data = result.get_data()
text = result.get_text()
meta = result.get_meta()

Use get_result() when the same model call will be inspected in more than one way.

2. Contract-First Output Control

Fixed required leaves belong directly in .output(...) as the third tuple item:

ticket = (
    agent
    .input("The billing export fails for accounts with archived invoices.")
    .output({
        "category": (str, "billing / auth / data / unknown", True),
        "severity": (int, "1-5", True),
        "next_actions": [(str, "recommended action")],
    })
    .start()
)

Use ensure_keys= for conditional or runtime-dependent paths. Use .validate(...) or validate_handler= for value-level business rules. Use ensure_all_keys=True when the whole schema must be present.

YAML and JSON prompt files can carry the same contract through $ensure: true, so teams can review prompt and response shape outside Python code.

3. Structured Streaming

Instant events let a UI, service, or downstream consumer react as each structured field changes:

result = (
    agent
    .input("Explain recursion with two examples.")
    .output({
        "definition": (str, "one sentence", True),
        "examples": [(str, "example with explanation")],
    })
    .get_result()
)

for event in result.get_generator(type="instant"):
    if event.path == "definition" and event.delta:
        print(event.delta, end="", flush=True)
    if event.wildcard_path == "examples[*]" and event.is_complete:
        print("\nEXAMPLE:", event.value)

This is useful for dashboards, chat UIs, SSE responses, and workflows that need partial structured results before the final response is complete.

4. Actions and Tool Use

Actions are model-callable capabilities. New code should start with @agent.action_func and agent.use_actions(...):

from agently import Agently

agent = Agently.create_agent()

@agent.action_func
def calculate_total(price: float, quantity: int) -> float:
    """Calculate an order total."""
    return price * quantity

agent.use_actions(calculate_total)

result = (
    agent
    .input("Use the available action to calculate 19.5 * 4, then explain the result.")
    .get_result()
)

print(result.get_text())
print(result.full_result_data["extra"].get("action_logs", []))

Common capability helpers:

agent.enable_python()
agent.enable_shell(root=".", commands=["pwd", "rg"])
agent.enable_workspace(root=".", read=True, write=False)
agent.enable_nodejs()
agent.enable_sqlite(database="app.db")

Built-in action packages:

from agently.builtins.actions import Browse, Search

agent.use_actions(Search(timeout=15, backend="duckduckgo"))
agent.use_actions(Browse())

Use agent.use_mcp(...) for MCP servers. Use agent.register_action(..., executor=..., execution_environments=[...]) when building a custom backend with explicit managed resources.

Instruction-heavy actions keep later model context compact with execution digests and artifact references. The application can read raw artifacts explicitly when it needs full code, shell output, page content, SQL rows, screenshots, or logs:

turn = agent.input("Use the shell action and summarize the result.")
records = agent.get_action_result(prompt=turn.prompt)
artifact_ref = records[0]["artifact_refs"][0]

raw = agent.action.read_action_artifact(
    artifact_id=artifact_ref["artifact_id"],
    action_call_id=artifact_ref["action_call_id"],
)

The older tool_func / use_tools / use_mcp / use_sandbox family remains a compatibility surface, but new examples use actions.

5. Runtime Skills

Skills are reusable task guidance and capability packages. In 4.1.3, the recommended application surface is agent.use_skills(...): declare candidate Skill sources on the Agent and let the Skills Executor perform lightweight discovery, planner selection, on-demand materialization, capability mounting, and execution diagnostics.

agent.use_skills(
    [
        {"source": "GarethManning/education-agent-skills"},
        {"source": "anthropics/skills", "subpath": "skills/docx"},
        {"source": "anthropics/skills", "subpath": "skills/pptx"},
        {"source": "anthropics/skills", "subpath": "skills/xlsx"},
    ],
    mode="model_decision",
)

execution = await agent.async_run_skills_task(
    "Create a four-week B1 business English course package.",
    effort="normal",
    output={
        "course_plan": (dict, "course goals and weekly structure", True),
        "teacher_guide": (str, "teacher-facing guide summary", True),
        "progress_tracker": ([str], "tracking columns and checkpoints", True),
    },
)

Business value: reusable Skills can turn a model call into a deliverable business process such as an education package, research memo, QA evidence pack, travel plan, or operational review. Application code stays focused on business inputs and output contracts instead of cloning remote repositories, parsing Skill files, or manually wiring each tool.

Skill-declared MCP, shell, and script capabilities mount through Action Runtime and Execution Environment, so side effects remain observable and policy controlled. High-risk local execution requires approval or auto_allow=True; safe pure-computation gaps can be synthesized as sandboxed Python actions, and business-system capabilities fail closed unless a real connector is mounted.

6. TriggerFlow Orchestration

TriggerFlow is Agently's workflow layer for explicit stages, branching, fan-out, event-driven input, runtime streams, pause/resume, persistence, and restart-safe execution.

import asyncio
from agently import TriggerFlow, TriggerFlowRuntimeData

flow = TriggerFlow(name="ticket-flow")

async def classify(data: TriggerFlowRuntimeData):
    text = data.input["text"]
    category = "billing" if "invoice" in text.lower() else "unknown"
    await data.async_set_state("category", category)
    return category

async def route(data: TriggerFlowRuntimeData):
    category = data.input
    await data.async_set_state("handler", f"{category}-team")

flow.to(classify).to(route)

async def main():
    execution = flow.create_execution()
    await execution.async_start({"text": "Invoice export failed."})
    snapshot = await execution.async_close()
    print(snapshot)

asyncio.run(main())

For services, workers, webhooks, human review, or SSE/WebSocket routes, keep the execution handle and close it explicitly:

execution = flow.create_execution(auto_close=False)
await execution.async_start(initial_input)
await execution.async_emit("UserApproved", {"approved": True})
snapshot = await execution.async_close()

close() / async_close() is the canonical completion path in the 4.1 line. The close snapshot is the durable result contract.

TriggerFlow is the right tool when you need:

7. Dynamic Task

Dynamic Task is Agently's framework-level surface for executing model-generated or app-generated DAGs. It is an application API, not a TriggerFlow sub-API; internally the executor compiles validated task graphs to TriggerFlow so it can reuse lifecycle, stream, pause/resume, and runtime resource mechanics.

from agently import Agently

async def local_handler(context):
    return {"task_id": context.task.id, "deps": dict(context.dependency_results)}

task = Agently.create_dynamic_task(
    target="review policy",
    plan={
        "graph_id": "review",
        "task_schema_version": "task_dag/v1",
        "tasks": [
            {"id": "extract", "kind": "local", "binding": "local_handler"},
            {"id": "final", "kind": "local", "binding": "local_handler", "depends_on": ["extract"]},
        ],
        "semantic_outputs": {"final": "final"},
    },
    handlers={"local_handler": local_handler},
)
snapshot = await task.async_start(timeout=10)

Use Dynamic Task when the task plan itself is data that needs planning, validation, pruning, and execution. Use TriggerFlow directly when you own a stable workflow topology in code.

8. Session Memory

Session keeps bounded multi-turn state when the problem is still one conversational thread, not a full workflow:

agent.activate_session(session_id="user-42")
agent.set_settings("session.max_length", 10000)

reply1 = agent.input("My name is Alice.").start()
reply2 = agent.input("What is my name?").start()

For long-running processes, event waits, fan-out, or human approvals, put TriggerFlow above the request layer instead of stretching Session into a workflow store.

9. Knowledge, Services, and Observability

Agently includes integration surfaces around the request and workflow layers:

• Knowledge base helpers for retrieval-backed context.
• FastAPIHelper for POST, SSE, and WebSocket exposure of agents, requests, generators, TriggerFlow definitions, and TriggerFlow executions.
• Observation events for request, action, execution environment, and TriggerFlow internals.
• Optional agently-devtools for local observation, evaluation, playground workflows, and project scaffolding.

pip install agently-devtools
agently-devtools init my_project

Agently 4.1.3.7 recommends agently-devtools >=0.1.9,<0.2.0.

Architecture

Layer Model

Agently organizes AI application code into explicit layers. The layers can be used independently, but they are designed to compose:

graph TB
    App["Application and business logic"]
    Settings["Settings files and environment"]
    Prompt["Prompt slots and output schema"]
    Agent["Agent request layer"]
    Model["Model requester plugins"]
    Result["ModelResponseResult: text, data, meta, stream"]
    Action["Action Runtime: planning, dispatch, logs"]
    Env["Execution Environment: MCP, Python, Bash, Node, Browser, SQLite"]
    Flow["TriggerFlow: branch, fan-out, stream, pause/resume, persist"]
    Observe["Observation events and DevTools"]

    App --> Settings
    App --> Agent
    Settings --> Agent
    Prompt --> Agent
    Agent --> Model
    Model --> Result
    Agent --> Action
    Action --> Env
    App --> Flow
    Flow --> Agent
    Agent --> Observe
    Action --> Observe
    Env --> Observe
    Flow --> Observe

Action Stack

Action Runtime separates planning, loop orchestration, backend execution, and managed resource lifecycle:

graph LR
    Agent["Agent"]
    Facade["Action facade"]
    Runtime["ActionRuntime plugin\nplanning and call normalization"]
    Flow["ActionFlow plugin\naction loop and orchestration bridge"]
    Executor["ActionExecutor plugin\nfunction, MCP, sandbox, Search/Browse, custom"]
    Env["ExecutionEnvironmentProvider\nresource lifecycle"]
    Logs["action_logs and artifacts"]

    Agent --> Facade
    Facade --> Runtime
    Runtime --> Flow
    Flow --> Env
    Flow --> Executor
    Executor --> Logs

Extension points:

Project Shape

For anything beyond a small script, keep settings, prompts, actions, flows, and service code separate:

my-agently-app/
  pyproject.toml
  .env
  settings.yaml
  prompts/
    summarize.yaml
    triage.yaml
  app/
    agents.py
    actions.py
    api.py
    main.py
  flows/
    triage.py
  tests/
    test_triage_flow.py

settings.yaml:

plugins:
  ModelRequester:
    OpenAICompatible:
      base_url: ${ENV.OPENAI_BASE_URL}
      api_key: ${ENV.OPENAI_API_KEY}
      model: ${ENV.OPENAI_MODEL}
debug: false

Load at startup:

from agently import Agently

Agently.load_settings("yaml_file", "settings.yaml", auto_load_env=True)

Prompt files can carry prompt slots and output contracts:

.request:
  instruct: |
    You are a concise editor. Keep facts intact.
  output:
    title:
      $type: str
      $ensure: true
    body:
      $type: str
      $ensure: true

Examples

Recommended model-app examples call a real model through DeepSeek or local Ollama and include an Expected key output source comment with stable key values from one real run.

Useful entry points:

Archived examples live under examples/archived/ and are compatibility references, not the recommended starting point for new apps.

Ecosystem

Agently Skills

Agently-Skills gives coding agents current Agently implementation guidance.

• Repository: https://github.com/AgentEra/Agently-Skills
• Current catalog generation: v2
• Recommended bundle: app
• Agently 4.1.3 compatibility: Skills authoring protocol agently-skills.authoring.v2

Use it when asking Codex, Claude Code, Cursor, or another coding agent to implement Agently patterns.

Agently DevTools

agently-devtools is optional and covers local observation, evaluation, interactive wrappers, and project scaffolding.

pip install agently-devtools
agently-devtools init my_project

Integrations

FAQ

What makes Agently different from direct SDK calls?

Direct SDK calls are excellent when the app only needs a small number of prompts. Agently adds contracts around those calls: prompt slots, output parsing, validation, retries, response reuse, action logs, session memory, configuration, service helpers, and TriggerFlow.

How is Agently different from LangChain?

LangChain provides broad integrations, prebuilt agents, and flexible building blocks. Agently is narrower and more opinionated about the model request boundary: provider setup, prompt slots, structured output, parser feedback, retries, validation, response reuse, action execution, settings, and observability are designed to line up as one contract. The goal is to let teams change the underlying model or provider without forcing downstream business logic to relearn output or tool-call shape.

How is Agently different from LangGraph?

LangGraph is strong when the central problem is graph-based agent state and durable execution. TriggerFlow is designed as Agently's signal-driven workflow layer: model-side instant structured events, action results, external events, pause/resume, runtime stream items, execution state, and close snapshots can all participate in the same orchestration story.

How is Agently different from CrewAI or AutoGen?

CrewAI and AutoGen are strong choices when the primary design is collaboration among agents. Agently is a lower-level application framework: multi-agent collaboration can be built as one pattern on top of structured model requests, Actions, TriggerFlow signals, sub-flows, Session, runtime resources, and service-facing execution handles.

Do I need TriggerFlow for every multi-step task?

No. Use plain Python or async functions for simple linear work. Use TriggerFlow when you need branches, fan-out, external events, pause/resume, runtime stream, persistence, or restart-safe execution.

Can I keep using older tool APIs?

Yes. The old tool family remains a compatibility surface and maps into the current action runtime. New code should prefer @agent.action_func, agent.use_actions(...), and the enable_* helpers.

How do I deploy an Agently service?

Use the async request APIs directly or wrap agents, requests, generators, TriggerFlow definitions, or TriggerFlow executions with FastAPIHelper. See the FastAPI docs and examples/fastapi/.

Documentation

Compatibility Notes

• The current package version is 4.1.3.7.
• The current release manifest is compatibility/releases/4.1.3.7.json.
• Development-line planning belongs in compatibility/in-development.json; do not treat planned future versions as released.
• README examples use the current Action and TriggerFlow close-snapshot paths.
• Deprecated APIs emit warnings once per Python process unless runtime.show_deprecation_warnings is disabled.

Community

• Discussions: https://github.com/AgentEra/Agently/discussions
• Issues: https://github.com/AgentEra/Agently/issues
• WeChat group: https://doc.weixin.qq.com/forms/AIoA8gcHAFMAScAhgZQABIlW6tV3l7QQf
• Twitter / X: https://x.com/AgentlyTech

License

Agently follows an open-core model:

• Open-source core: Apache 2.0
• Trademark usage policy: TRADEMARK.md
• Contributor rights agreement: CLA.md
• Enterprise extensions and services: separate commercial agreements