Deep Insight splits its work across Coordinator, Planner, Plan Reviewer, Supervisor, Coder, Validator, Reporter and Tracker agents. The Strands Agents SDK Graph pattern sets the execution order, and the Agents-as-Tools pattern lets the Supervisor call specialized agents only when a step needs them. A human-in-the-loop plan review can send the plan back to the Planner for revision.
Each role uses a different Claude model: Haiku for routing, Opus with extended thinking for planning, and Sonnet for execution. Prompt caching is enabled for the Supervisor, Coder and Reporter, which the post says can reduce input token costs by up to 90 percent on repeated calls. Generated Python runs in per-session containers on AWS Fargate behind an Application Load Balancer, inside a private VPC, and containers terminate after one hour.
The system ships in three versions: a Self-Hosted CLI for development, a browser-based Web version, and a Managed Amazon Bedrock AgentCore version for session isolation and private VPC deployment. An optional Ops dashboard records task metrics in DynamoDB, sends failure alerts through Amazon SNS, and uses Amazon Cognito for access. The post recommends adding Amazon Bedrock Guardrails for content filtering and grounding checks.
Why it matters if you build agents
Teams building multi-agent systems can reuse the Graph and Agents-as-Tools patterns in the Strands Agents SDK, along with the approach of matching a model to each agent's role to balance cost and quality. The per-session Fargate sandbox is a concrete pattern for running LLM-generated code without exposing a production server. The post's point that prompt changes alone did not fix business-logic and validation failures suggests building explicit verification and approval steps into the workflow.
Get stories like this every weekday
One short email with the day's agent tooling news.