Agently 4.1.3.7 Release Notes

Languages: English · 中文

Agently 4.1.3.7 hardens the AgentExecution-backed AgentTaskLoop path. The release keeps AgentExecution as the recommended Agent run owner while making goal pursuit clearer, easier to validate, and better aligned with Skills, Actions, DAG-shaped bounded steps, Workspace evidence, and release diagnostics.

This is not a new public AgentTask lifecycle and not a TriggerFlow distributed recovery milestone. Future AdaptiveLoop / BootstrapLoop packaging, multi-task scheduling, background autonomy, and resumable AgentExecution strategies remain explicitly deferred.

Recommended Shape

Use one execution draft and configure Prompt, Skills, Actions, goals, and effort around that draft:

result = (
    agent
    .use_skills("website-builder", "seo-reviewer")
    .use_actions(write_file, read_file)
    .goal(
        [
            "Build a small product website.",
            "Include brand introduction, product features, target users, and contact information.",
        ],
        success_criteria=[
            "The final artifact is a runnable page file.",
            "The page content covers every supplied business fact.",
            "Execution evidence includes file write, readback, and content inspection.",
        ],
    )
    .effort(
        "medium",
        budget={"iteration_limit": 3, "model_call_limit": 10},
        execution={"step_plan": "auto"},
        verification={"strictness": "normal"},
    )
    .start()
)

data = result.get_data()
task_refs = result.task_refs
meta = result.get_meta()

agent.goal(goal_or_goals, success_criteria=None) and the plural readability alias agent.goals(...) return task-strategy AgentExecution drafts. Detailed effort(...) controls planning depth, budget, verification, replan, progress, and optional DAG-shaped bounded steps; it does not grant permissions or bypass host policy.

Core Changes

Area	What changed	Recommended usage	Compatibility / risk	Evidence
Goal Pursuit API	Goal and success criteria are one execution-first configuration surface.	agent.goal([...], success_criteria=[...]).effort(...).start()	Additive recommended shape; AgentTask is still not the primary public lifecycle.	AgentTaskLoop tests, release manifest, docs, examples.
Effort and bounded steps	effort(...) now carries budget, planning, verification, replan, progress, and execution.step_plan guidance.	Use presets for normal tasks; add detailed sections only when the host needs explicit bounds.	execution.step_plan guides strategy only; DAG completion is step evidence, not task acceptance.	tests/test_agent_task_loop.py, auto-orchestration docs, compatibility manifest.
Verification and replan	Goal Pursuit reports configured, planned, executing, evidence, verified, guarded, replan, and terminal phase concepts.	Consume result, stream, task refs, and metadata through AgentExecutionResult.	Completion still requires model verifier acceptance plus host guards.	AgentTaskLoop tests and examples/agent_task/goal_pursuit_acceptance_matrix.py.
Skills context packs	SkillsExecutor exposes task-aware context-pack and TaskDAG resolver support for planner and DAG consumption.	Use agent.use_skills(...); custom planners can call context-pack APIs when needed.	SkillsExecutor remains a capability adapter and does not execute bundled scripts directly.	SkillsExecutor tests, main compatibility metadata, Agently-Skills validation.
ActionRuntime blockers	Action metadata redacts execution-environment secrets, action-loop timeout covers the full loop, explicit action-loop replies do not re-enter ActionRuntime, and empty native tool-call plans return diagnostics.	Use current Action APIs and inspect agent.action.summarize_records(...) for host-facing records.	Fixes release blockers without changing ActionRuntime ownership.	ActionRuntime tests and compatibility manifest.
DevTools companion	DevTools UI, playground, and runtime observation semantics align with agent_execution naming.	Release Agently 4.1.3.7 with agently-devtools 0.1.9 and keep the main manifest recommendation at >=0.1.9,<0.2.0.	Synchronous companion release required; do not publish main with a stale DevTools recommendation.	Main compatibility manifest, DevTools package metadata, DevTools tests.
TriggerFlow and Workspace foundation	Current manifests record the landed TriggerFlow runtime integrity and Workspace durable-provider contracts that support this release line.	Treat TriggerFlow recovery and Workspace provider seams as foundation capabilities; do not present future distributed recovery as complete AgentTask work.	Deferred production distributed recovery remains explicit.	TriggerFlow / Workspace tests, examples, and specs.

Compatibility

• Package version: 4.1.3.7.
• Release manifest: compatibility/releases/4.1.3.7.json.
• Recommended agently-devtools: >=0.1.9,<0.2.0.
• AgentTurn, create_turn(...), set_turn_prompt(...), set_request_prompt(...), one_turn, task_step, and task_scope are not current 4.1.3.7 recommended surfaces.
• agent.resume(task_id) / await agent.async_resume(task_id) resume a checkpointed AgentTaskLoop as a task-strategy AgentExecution. AgentExecutionResult.resume() delegates to the same Agent facade when the result carries resumable task_refs; resume_task(...) remains a compatibility alias, not the recommended lifecycle.

Validation Summary

• Static typing and deterministic tests cover AgentExecution chaining, AgentTaskLoop verification/replan, progress and snapshot streams, task refs, Workspace evidence links, Skills context-pack integration, ActionRuntime release blockers, TriggerFlow runtime/resource behavior, Workspace provider contracts, and compatibility registry alignment.
• The Goal Pursuit release example examples/agent_task/goal_pursuit_acceptance_matrix.py records a real local Ollama run on 2026-06-12 with an accepted task and a max-iteration partial task. Key output includes accepted.status="completed", accepted.artifact_status="accepted", partial.status="max_iterations", partial.accepted=false, and partial.guard_reasons containing missing_criteria_present.
• Agently-Skills guidance was updated for the same 4.1.3.7 usage shape and should be validated with the companion validation suite before promotion.

Deferred Scope

4.1.3.7 deliberately does not complete multi-task scheduling, background autonomy, distributed lease ownership, full durable pause/resume, production external storage providers, or TriggerFlow-backed AdaptiveLoop / BootstrapLoop packaging for AgentTaskLoop. Those remain future architecture work.