Agent Run
Deploy. Monitor. Scale.
Agent Run executes published agent definitions for a user, an application, a scheduled job or an incoming event, with a choice of two frameworks: LangGraph / LangChain or Strands Agents.
More than a model call
A production runtime manages much more than calls to a language model. Agent Run handles execution identity, conversation state, tool calls, progress events, failures, results and operational visibility for every run.
Option A: LangGraph / LangChain
| Technology | Role |
|---|---|
| LangChain | Agent abstractions and integrations for models, tools, prompts and middleware |
| LangGraph | Stateful orchestration of workflow steps, agent loops, branching and execution state |
LangGraph mixes deterministic application logic with model-driven decisions. LangChain agents are built on LangGraph, and LangGraph can also be used on its own. This option suits teams that want to define and inspect the structure of a business process, for example:
- Retrieve a request and its supporting documents.
- Extract and validate the required information.
- Route incomplete cases for clarification.
- Perform the analysis.
- Pause for approval.
- Produce the final result.
LangGraph checkpoints keep execution state for each conversation thread, with separate stores for longer-term application data. Its interrupt mechanism pauses a run for human input and resumes it with the supplied response, while Agent Run provides the approval interface and authorization rules around it.
Option B: Strands Agents
Strands provides an agent loop that repeatedly invokes a model, runs the tools it requests and returns the results, until the agent responds or reaches an execution limit. The model decides which available action to take next. For example, a research assistant receives a question, identifies what information is missing, chooses relevant retrieval tools and synthesizes the findings.
Strands also supports several ways to coordinate agents:
| Pattern | How it organizes work |
|---|---|
| Agents as tools | An orchestrating agent calls specialist agents |
| Graph | Nodes and edges define dependencies and execution paths, including conditional edges, parallel paths and cycles |
| Swarm | Agents hand work to other specialists |
| Workflow | Tasks run through a defined dependency structure |
Session persistence and observability support for traces, metrics and logs connect to the platform's storage and monitoring services.
Choosing a framework
| Customer priority | LangGraph / LangChain | Strands Agents |
|---|---|---|
| Explicit process orchestration | State, workflow structure and controlled transitions | Graph or workflow patterns where appropriate |
| Adaptive tool use | Agent behavior configured within the orchestration | Model-directed agent loop |
| Specialist collaboration | Agents composed within workflows | Agents-as-tools, graph and swarm patterns |
| Conversation continuity | Configured checkpoints and stores | Configured session and state persistence |
| Framework preference | Teams using the LangChain ecosystem | Teams using the Strands ecosystem |
Neither framework is limited to these use cases. The table shows where each one is strongest.
Shared platform capabilities
Whichever framework you choose, Agent Run provides:
- A consistent interface for starting and inspecting runs.
- Run history, status and streamed progress.
- User and workspace permissions.
- Tool credential management.
- Execution limits, cancellation and error reporting.
- Usage tracking and evaluation.
- Framework-specific state persistence and recovery.
A shared interface does not make workflows or saved state automatically portable between frameworks. Moving a workflow from one framework to the other requires validation, and tools that change external systems are designed so that a retried step does not repeat its action.
Framework choice within a shared product experience.