flyte.ai.agents.agent

Agent — a flyte-native tool-use agent harness.

The harness operates a simple, robust LLM tool-call loop:

1. Send the conversation + tool catalog to the LLM.
2. If the assistant returns tool calls, execute each one (sequentially or concurrently), append the results back into the message history, and loop.
3. Stop when the assistant returns a plain text reply (no tool calls) or when max_turns is reached.

Design goals

• Minimal core: a single agent loop with clear extension points.
• Tools = anything callable: plain functions, @flyte.trace helpers, @env.task :class:~flyte.TaskTemplate instances (durable, on-cluster execution), LazyEntity references to remote tasks, or pre-built :class:AgentTool instances.
• MCP integration: connect to external MCP servers (Slack, GitHub, Linear, filesystem, etc.) and surface their tools to the agent transparently.
• Memory: optional :class:MemoryStore that serializes conversation history plus path-addressed artifacts to / from :class:flyte.io.Dir, letting the agent persist state across runs (e.g. across scheduled wake-ups or webhook invocations). Includes opt-in audit log, read-only prefixes, and optimistic concurrency for multi-agent / sleep-wake patterns.
• HITL: opt-in per-tool human approval that pauses the loop and waits for a human via the flyteplugins-hitl plugin before executing the tool.
• Flyte-native: implements the :class:AgentProtocol so it works seamlessly with :class:~flyte.ai.chat.AgentChatAppEnvironment and is happy to be wrapped in @env.task(triggers=...) for scheduled or webhook-driven wake-ups.

Heavy inspiration is taken from the pi <https://github.com/earendil-works/pi>_ agent harness — in particular its event model and the separation of the loop from message conversion / tool invocation.

Implementation note: the tool/MCP/LLM building blocks live in sibling modules (:mod:._tools, :mod:._mcp, :mod:._llm); this module focuses on the Agent class and the loop.

Directory

Classes

Variables