Haystack

Agents call tools built from any component, pipeline or Python function, MCP tools and toolsets, and ready-made tools such as GitHub file editing and issue commenting.

SourceHaystack, docs.haystack.deepset.ai tools pagesread 2026-09-21

Pipelines connect modular components into graphs with branches, loops and cycles, agents run tool-calling loops inside them, and multi-agent systems let a coordinator agent delegate to specialist agents, so deterministic steps and agent steps combine in one run.

SourceHaystack, docs.haystack.deepset.ai concepts/agents/multi-agent-systems and haystack.deepset.airead 2026-09-21

Indexing pipelines write the customer's documents into document stores such as Weaviate, Pinecone and Elasticsearch, queried by keyword, embedding and hybrid retrievers with rankers and metadata filters, and the Enterprise Platform adds data indexing in its builder, which is a maintained retrieval structure over the customer's knowledge.

SourceHaystack, docs.haystack.deepset.ai retrievers and document stores, and deepset.ai/haystack-platformread 2026-09-21

A confirmation hook on the agent intercepts tool calls before execution so a person can confirm, reject or modify each one, with policies to always ask, ask once or never ask, and the Enterprise Platform adds guardrails and policy enforcement, so autonomy can differ between low and high risk actions.

SourceHaystack, docs.haystack.deepset.ai pipeline-components/agents/human-in-the-loop and deepset.ai/haystack-platformread 2026-09-21

The Haystack Enterprise Platform documents role-based access control, single sign-on and audit logs, and deepset states it is SOC 2 Type II and ISO 27001 certified, GDPR and HIPAA compliant and CSA STAR Level 1.

SourceHaystack, deepset.ai/haystack-platformread 2026-09-21

The Enterprise Platform keeps a unified run history of every query, answer and context, with traces, component-level error monitoring, usage reports and audit logs, and the open source framework exports traces to OpenTelemetry, Datadog, Langfuse, MLflow and a logging tracer.

SourceHaystack, deepset.ai/haystack-platform and docs.haystack.deepset.ai development/tracingread 2026-09-21

Haystack's own memory is run state: a State container shares data across tool calls within an agent run, and pipeline breakpoints save snapshots to resume from. Long-term memory comes through memory stores backed by Mem0's cloud API and Cognee, products sold separately, and no memory layer of Haystack's own with a stated scope and lifetime is documented.

SourceHaystack, docs.haystack.deepset.ai agents state, memory-stores and pipeline-breakpointsread 2026-09-21

The Enterprise Platform deploys to managed cloud or self-hosted on the customer's own infrastructure, described as sovereign and on-premises capable, and the open source framework ships deployment guides for Docker, Kubernetes and OpenShift.

SourceHaystack, deepset.ai/haystack-platform and docs.haystack.deepset.ai development/deploymentread 2026-09-21

Agent Pack is a collection of complex, pre-configured agents to run as they are, customize or copy as a blueprint, including an advanced RAG agent and a deep research agent, and the Enterprise Platform starts builds from pipeline and agent templates; each agent does its own job and stands alone without the other.

SourceHaystack, docs.haystack.deepset.ai agents/agent-pack and deepset.ai/haystack-platformread 2026-09-21

Pipelines and agents run when called: Hayhooks serves them through HTTP endpoints and the Enterprise Platform deploys them behind a REST API, so any outside system can be wired to start one, but every source is custom-wired rather than a built-in channel, and no schedule or event source that wakes an agent on its own is documented.

SourceHaystack, docs.haystack.deepset.ai development/hayhooks and deepset.ai/haystack-platformread 2026-09-21

The developer picks generators from integrations with OpenAI, Anthropic, Mistral, Hugging Face and others with no vendor lock-in, and deepset describes a customer swapping LLMs without breaking the production system, so model choice sits with the customer.

SourceHaystack, haystack.deepset.ai and deepset.ai/haystack-platformread 2026-09-21

Haystack is a documented Python SDK installed with pip, Hayhooks serves pipelines and agents through HTTP endpoints, and the Enterprise Platform exposes deployed pipelines through a REST API and exports them as Python or YAML, so outside callers drive the platform.

SourceHaystack, docs.haystack.deepset.ai development/hayhooks, haystack.deepset.ai and deepset.ai/haystack-platformread 2026-09-21

Haystack evaluates entire pipelines or single components such as retrievers and generators with statistical and model-based evaluators, including exact match, context relevance and document ranking metrics, and the Enterprise Platform compares prompts, retrieval strategies and agent workflows side by side before production.

SourceHaystack, docs.haystack.deepset.ai optimization/evaluation and deepset.ai/haystack-platformread 2026-09-21

Neither Haystack nor the Enterprise Platform documents browser, desktop or computer control: Haystack's full documentation text carries no browser automation, headless browser or computer use capability, and its ready-made tools are API and code-execution tools rather than interface control.

SourceHaystack, docs.haystack.deepset.airead 2026-09-21