Agently 4.1.3.6 Release Notes

Languages: English · 中文

Agently 4.1.3.6 is the AgentExecution ownership release. It consolidates the recommended public surface around AgentExecution and AgentExecutionResult, hardens OpenAI-compatible stream completion, and publishes a narrow single-task task-loop slice without presenting it as the full future AgentTask system.

What Changed

AgentExecution is the one-run owner

Quick prompt chains now align with the same public ownership model as explicit executions:

execution = (
    agent
    .input("Classify this customer request.")
    .output({"category": (str, "billing, support, or sales", True)})
)

result = execution.get_result()
data = result.get_data()
meta = result.get_meta()

AgentTurn, create_turn(...), and set_turn_prompt(...) remain compatibility surfaces for older 4.1.3 examples and migration paths, but they are no longer the recommended public lifecycle. The request-local prompt draft, route choice, stream, metadata, and result facade now belong to AgentExecution.

Result-first consumption is the recommended shape

Use get_result() when the same run may be read as text, data, metadata, stream events, or task refs. Agent quick prompt chains return AgentExecutionResult; direct ModelRequest builders return ModelResponseResult.

Result-named stream aliases are now the recommended root-level names. Older response-named aliases remain compatibility-only under agently.types.data and are deprecated for Agently 4.2 removal.

Bounded task loops are carried by AgentExecution

agent.create_task(...) and agent.create_task_loop(...) return task-strategy AgentExecution drafts. The current slice is intentionally narrow:

• one business task
• one Agent owner
• bounded iteration guidance, roughly 2-5 rounds
• explicitly enabled Actions, Skills, or Dynamic Task candidates
• model-owned planning, verification, and replan
• conservative host guards for missing criteria, risky action evidence, approval-required actions, and final deliverables

execution = agent.create_task(
    "Prepare a customer-safe incident update from the provided evidence.",
    success_criteria=[
        "Names the customer impact",
        "Separates confirmed facts from unknowns",
        "Lists the next customer-facing action",
    ],
    max_iterations=3,
)

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

completed means the model verification accepted the result and host guards accepted the artifact. Reaching max_iterations can still return accepted=false and artifact_status=partial. AgentExecutionResult.resume() is reserved and reports supported=false until a resumable strategy lands.

OpenAI-compatible stream completion is hardened

Issue #287 is fixed on the 4.1.3.6 line. Some OpenAI-compatible gateways send a usage-only final SSE chunk with missing or empty choices before [DONE]. OpenAICompatible.broadcast_response(...) now preserves the accumulated content and synthesizes the terminal message instead of indexing into an empty list.

The same stream-end path now treats GeneratorExit as control flow instead of a model requester error, avoiding a spurious empty model.requester.error after an otherwise successful stream.

Release guardrails now include foundation examples

The release workflow now requires a Foundation example effect gate for substrate capabilities touched or claimed by a release, such as ModelRequest/ModelResponse, TriggerFlow, Dynamic Task/TaskDAG, ActionRuntime, ExecutionEnvironment, and provider protocols. Tests are not enough on their own: the corresponding core examples/ scenario must be run against the release candidate, using real DeepSeek or local Ollama when model-owned behavior is involved.

Compatibility

• Package version: 4.1.3.6.
• Release manifest: compatibility/releases/4.1.3.6.json.
• Recommended agently-devtools: >=0.1.8,<0.2.0.
• AgentTurn, create_turn(...), set_turn_prompt(...), and set_request_prompt(...) remain compatible migration surfaces.
• DevTools display issues #288 and #289 are DevTools-side work and are not closed by this Agently package release.

Validation Summary

• Static typing and tests cover AgentExecution result facades, task-loop result refs, stream/meta access, OpenAI-compatible usage-only final chunks, and compatibility registry alignment.
• Foundation example effect gate for the touched ModelRequest/ModelResponse substrate passed with DeepSeek: python examples/step_by_step/05-response_result.py produced result_type=ModelResponseResult, data_has_definition=True, meta_has_id=True, result_cached=True, and delta_event_count_positive=True.
• The 4.1.3.6 AgentExecution use-case example examples/agent_auto_orchestration/22_unified_agent_execution_result.py passed with DeepSeek and produced quick_result_type=AgentExecutionResult, quick_category=renewal_risk, task_strategy=task_loop, and task_result_status=completed.
• Agently-Skills guidance was updated with the same foundation example gate and validated with the companion validation suite.

Deferred Scope

This is not the complete future AgentTask system. Multi-task scheduling, distributed leases, background autonomy, full durable pause/resume, and Workspace-as-automatic-memory planning remain deferred. The 4.1.3.7 line should harden the AgentExecution-carried task-loop strategy without replacing the public ownership boundary introduced here.