Agently 4.1.3.3 Release Notes

Languages: English · 中文

Agently 4.1.3.3 is a release-line hardening slice for the 4.1.4 AgentTask target. It closes the public configuration issues in #274 and #276, makes model request retries and runtime event ownership easier to reason about, and adds a small .image(...) convenience API for common VLM requests.

What Changed

• agent.create_execution(...) now accepts dict-compatible typed options=.... The first routed consumer is routes.skills.effort, so Skills auto-orchestration can use the same fast / normal / max effort control as explicit Skills calls.
• Model routing now supports the recommended layered shape: model_pool -> model_profiles -> api_key_pools. A business model key can resolve to a provider profile with provider name, model, base URL, request/client options, and an API key pool.
• api_key_pools now separates request-time key selection from provider-error failover. Selection and failover both support built-in policies and custom handlers.
• Model requester plugins now cooperate with core through handler contracts. Official runtime events for framework-owned flows are emitted by core, while provider plugins return observations, errors, and decisions for core to map.
• The built-in model requester packages and agently/core layout were reorganized into package directories with stable public exports. Existing public imports continue to work.
• agent.image(...) and request.image(...) now build VLM image input from a question plus local files or remote URLs. Local PNG, JPEG, WebP, GIF, and BMP files are converted to data:<mime>;base64,... image URLs.
• output_format="instant" documentation now clearly describes immediate field streaming, its value, and its relationship to structured output modes.

Usage Shapes

Skills route effort through execution options:

from agently.types.options import ExecutionOptions, SkillsRouteOptions

execution = agent.create_execution(
    options=ExecutionOptions(
        routes={"skills": SkillsRouteOptions(effort="normal")},
    ),
)

Layered model routing:

agent.set_settings("model_pool", {"skills.reason": "deepseek.reasoner"})
agent.set_settings("model_profiles", {
    "deepseek.reasoner": {
        "provider": "OpenAICompatible",
        "base_url": "https://api.deepseek.com/v1",
        "model": "deepseek-reasoner",
        "api_key_pool": "deepseek.prod",
    },
})
agent.set_settings("api_key_pools", {
    "deepseek.prod": {
        "selection": {"strategy": "round_robin"},
        "failover": {"strategy": "try_next", "retry_status_codes": [429]},
        "keys": [
            {"id": "primary", "value": "${ENV.DEEPSEEK_API_KEY}"},
            {"id": "secondary", "value": "${ENV.DEEPSEEK_API_KEY_2}"},
        ],
    },
})

VLM image input:

result = (
    agent
    .image(
        question="Compare these two screenshots and list the visible differences.",
        files=["./before.png", "./after.png"],
    )
    .start()
)

Compatibility

• Package version: 4.1.3.3.
• Release manifest: compatibility/releases/4.1.3.3.json.
• Recommended agently-devtools: >=0.1.6,<0.2.0.
• Existing dict-based settings, legacy model_pool, key_pool_strategy, and key_pool shapes remain compatible.
• .attachment([...]) remains the low-level rich-content input surface. The new .image(...) API is convenience syntax for question plus image sources.

Issue Scope

This release marks #274 and #276 as solved on the development line. It prepares the configuration and request/runtime substrate for 4.1.4 AgentTask V1, but it does not implement AgentTask itself.