Haystack

Everything is a component with typed inputs and outputs. A Pipeline connects components into a directed graph; a component reruns whenever all its required inputs arrive again, which is how loops and self-correction work, and routers add branching. The same Pipeline class runs synchronously, asynchronously or as a stream, and pipelines serialize to YAML.

The Agent component wraps a chat generator and tools and loops until an exit condition. For multiple agents, the documented pattern is a coordinator agent that calls specialist agents wrapped with AgentTool; the coordinator sees only each specialist's final reply. Agents can also sit inside larger pipelines.

Within one run, State holds typed values (messages, documents and custom fields) that tools read from and write to. Longer-term memory comes from integrations such as Mem0 memory stores and document stores, and compaction hooks trim long conversations. Pipeline breakpoints capture snapshots that can be edited and resumed.

A ConfirmationHook registered under the Agent's before_tool hook point intercepts tool calls. A strategy (the built-in BlockingConfirmationStrategy) asks a person whether to confirm, reject or modify the call, a policy (AlwaysAskPolicy, AskOncePolicy, NeverAskPolicy) decides when to ask, and the built-in UIs are console prompts. Rejections are fed back to the model. Pipelines can also be paused with breakpoints and resumed from an edited snapshot; since 3.0, pausing inside an Agent is no longer supported.

pip install haystack-ai

from haystack.components.agents import Agent
from haystack.components.generators.chat import OpenAIChatGenerator
from haystack.dataclasses import ChatMessage
from haystack.tools import AgentTool, tool

@tool
def word_count(text: str) -> int:
    """Count the words in a piece of text."""
    return len(text.split())

editor = Agent(chat_generator=OpenAIChatGenerator(model="gpt-5.4-mini"), tools=[word_count],
               system_prompt="Tighten drafts and report their word count.")
lead = Agent(chat_generator=OpenAIChatGenerator(model="gpt-5.4-mini"),
             tools=[AgentTool(agent=editor, name="editor", description="Shortens a draft and counts words")],
             system_prompt="Delegate editing to the editor, then summarize the result.")

result = lead.run(messages=[ChatMessage.from_user("Shorten: 'We are very, very happy to announce the release.'")])
print(result["last_message"].text)

• The package is haystack-ai; Haystack 1.x (farm-haystack) is end of life.
• OpenAIChatGenerator reads OPENAI_API_KEY from the environment.
• Optional features raise an ImportError that names the missing dependency to install (for example pypdf).
• MCP tools need the separate mcp-haystack package.
• 3.0 removed ToolInvoker, AsyncPipeline and the legacy non-chat generators, and moved HITL to haystack.hooks.human_in_the_loop; follow the migration guide.
• Anonymous telemetry is on by default; set HAYSTACK_TELEMETRY_ENABLED=False to opt out.

Yes. MCPTool and MCPToolset from the mcp-haystack integration connect agents to MCP servers over Streamable HTTP or stdio, and Hayhooks can serve Haystack pipelines and agents as an MCP server. A2A and AG-UI support were not found.

The framework is Apache-2.0. deepset also sells the Haystack Enterprise Platform (managed or self-hosted) and Haystack Enterprise Starter support, with a free Studio tier and custom enterprise pricing on its pricing page.

A coordinator Agent receives specialist Agents wrapped with AgentTool as tools; it sends each specialist a task and gets back only the final reply.

Yes. A ConfirmationHook on the before_tool hook point lets a person confirm, reject or modify each tool call, with policies that decide when to ask.

The Agent gained a hooks system and owns tool execution; ToolInvoker, AsyncPipeline and legacy generators were removed, HITL moved to haystack.hooks.human_in_the_loop, and agent-level breakpoints were dropped.