Amazon Bedrock Agents
also known as Agents for Amazon Bedrock
Type: full-code · Vendor: Amazon Web Services · Language: API (multi-language SDKs: Python, Java, JavaScript, .NET, Go via AWS SDKs) · License: proprietary · Status: active · Status in practice: mature · First released: 2023-11-28
AWS-managed agent runtime that turns a Bedrock foundation model into a tool-using, knowledge-base-grounded agent: action groups (OpenAPI/function-detail schemas backed by Lambda or return-of-control), Bedrock Knowledge Bases for RAG, code interpretation, multi-agent supervisor/collaborator hierarchies, long-term memory, and guardrails.
Description. Amazon Bedrock Agents is a fully managed runtime for autonomous agents on AWS. At build time you assemble a foundation model, natural-language instructions, one or more action groups (each defined by an OpenAPI or function-detail schema and resolved via a Lambda function or by returning control to the developer), and optional Bedrock Knowledge Bases. At runtime the InvokeAgent API drives a pre-processing -> orchestration -> post-processing sequence where the model generates a rationale, picks an action or knowledge-base query, observes the result, and iterates until done. Multi-agent collaboration designates a supervisor agent that routes work to collaborator agents. Memory retains conversational context across sessions. Bedrock Guardrails apply content filters, denied topics, PII redaction, and contextual grounding checks.
Agent loop shape. Three-phase ReAct-style sequence per InvokeAgent request: (1) Pre-processing validates the user input; (2) Orchestration is a model-driven loop where the FM emits a rationale, predicts which action group to invoke or which knowledge base to query, the action runs (via Lambda or by returning control to the developer) or the KB is queried, the observation augments the base prompt, and the loop iterates; (3) Post-processing formats the final response. Conversation history persists per session, long-term memory summarises completed sessions per memoryId. Multi-agent collaboration wraps this loop with a supervisor that delegates to collaborator agents.
Primary use cases
- OpenAPI / function-detail action groups backed by Lambda
- RAG-grounded enterprise assistants via Bedrock Knowledge Bases
- supervisor-collaborator multi-agent workflows
- code-interpreter data analysis and file processing
- return-of-control for human or developer-side action execution
Key concepts
- Action groups → tool-use (docs) — OpenAPI or function-detail schemas backed by a Lambda function (or return-of-control) that define the actions an agent can take.
- Knowledge Bases → agentic-rag (docs) — Vector-backed RAG sources the agent queries to augment generation.
- Orchestration loop (docs) — Pre-processing -> rationale -> action / KB -> observation -> iterate, optional post-processing.
- Return of control → approval-queue (docs) — Action groups can return parameters to the caller instead of invoking Lambda, letting the developer execute and feed results back via invocationId.
- Multi-agent collaboration → supervisor (docs) — Designate a supervisor agent and associate collaborator agents; supervisor builds and executes a plan across them.
- Memory → cross-session-memory (docs) — Long-term memory summarises sessions per memoryId; retained up to 365 days.
- Code Interpretation → code-execution (docs) — Sandboxed code execution for data analysis, charting, and file processing.
Patterns this full-code implements —
- ★★Agent Resumption
Session conversation history persists across InvokeAgent calls per sessionId; memoryId summaries restore long-term context across new sessions.
- ★★Agentic RAG
Knowledge Bases are associated with an agent; the orchestration step explicitly chooses between action invocation and KB query.
- ★★Approval Queue
Return of Control routes action parameters back to the developer's application instead of Lambda, enabling explicit external execution / approval gating.
- ★★Code Execution
Code interpretation generates and runs code in a secure sandbox; supports analysis and file processing.
- ★★Cross-Session Memory
Memory feature retains context across multiple sessions per memoryId with configurable retention up to 365 days.
- ★★Orchestrator-Workers
Multi-agent collaboration breaks tasks across specialist sub-agents under a supervisor; collaborators can also work in parallel.
- ★★Supervisor
Explicit supervisor agent designation; collaborators are associated with the supervisor and routed to by it.
- ★★Tool Use
Action groups are the headline tool-use primitive: OpenAPI or function-detail schemas executed via Lambda or return-of-control.
- ★★Input/Output Guardrails
Bedrock Guardrails provide content filters, denied topics, PII redaction, and contextual grounding checks; applicable to agent inputs and outputs.
- ★★Structured Output
Action groups use OpenAPI / function-detail schemas to elicit typed parameters from the model; return-of-control surfaces those typed parameters as structured invocationInputs in the InvokeAgent resp…
- ★★Model Context Protocol
Bedrock Agents themselves use action groups (OpenAPI / Lambda), not MCP. MCP support is provided in the adjacent Amazon Bedrock AgentCore platform: AgentCore Gateway implements MCP and converts APIs,…
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Alternatives & relatives
Listed as alternative by (2)
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: verified