Agently.tech

Agently is an open-source Python framework from AgentEra for building AI service backends. It combines structured model outputs, controlled Actions, MCP support, runtime Skills, streaming, and recoverable workflows. The framework requires Python 3.10+, installs with pip, and its README lists current framework version 4.1.4.8.

What It Is

Agently is an AI application runtime framework aimed at product engineers and platform teams who need model behavior to fit stable application contracts. A request is composed from prompt slots such as role, info, instruct, input, and output. Results can be read as parsed data, text, or metadata. The output schema supports required fields, retries, and validation handlers, so downstream code receives typed fields rather than free text.

Structured Output and Streaming

The .output(...) schema defines fields the model must return. In instant streaming mode, consumers can react to individual fields while the model is still generating, which suits UI patches, SSE routes, and workflow signals. Provider setup is normalized across OpenAI-compatible endpoints (including DeepSeek, Ollama, and Qwen) and Anthropic-compatible endpoints, so switching models happens through settings. A Model Pool and API key pools support multiple models and failover.

Actions, Skills, and MCP

Actions are model-callable capabilities. They can be local functions registered with @agent.action_func, MCP servers via agent.use_mcp(...), Search/Browse packages, or helpers for Python, shell, Node.js, SQLite, and workspace files. Action results carry logs and artifact references. Runtime Skills are reusable guidance packages bound to an execution, and skill requirements do not grant side-effect permissions on their own.

Workflows with TriggerFlow and Dynamic Task

TriggerFlow is the event-driven workflow layer, with branching, fan-out, runtime streams, pause/resume, save/load, and close snapshots. Dynamic Task turns model-generated or application-submitted DAG data into validated, observable task execution. FastAPIHelper exposes agents and flows through POST, SSE, and WebSocket endpoints.

Companion Tooling

Agently-Skills is a package of installable skills that guides coding agents such as Codex, Claude Code, and Cursor in generating Agently code. The optional agently-devtools package provides local observation, evaluation, a playground, and project scaffolding. The open-source core uses Apache 2.0, and the project describes an open-core model with separate commercial agreements for enterprise extensions.