AWS Builder Center

Fuse: Building a Runtime Circuit Breaker for Autonomous AI Agents with AWS

Fuse — Circuit Breaker for Autonomous AI Agents Autonomous agents can loop, retry failed tools, and consume resources after useful progress stops. Fuse evaluates each model or tool invocation before execution. When limits for steps, runtime, estimated cost, or repeated actions are exceeded, Fuse blocks the next call and preserves an ordered audit trail. Safe workflows still complete normally. Built for developers creating reliable AI systems, Fuse turns unchecked autonomy into explainable execution.

Autonomous AI agents are powerful because they can reason, choose actions, and interact with tools independently. However, that autonomy also creates a serious operational challenge: an agent may repeat the same action, retry a failing tool indefinitely, or continue making model and tool calls after useful progress has stopped.
These behaviors are difficult to detect and even harder to control once execution has begun. A single ambiguous response can cause an agent to loop, increasing estimated usage and making the system’s behavior difficult to explain. Most applications focus on what an agent can do. Fuse focuses on when an agent should stop. Fuse is a runtime safety layer that acts as a circuit breaker for autonomous AI workflows. Before every integrated model or tool invocation, Fuse evaluates the proposed action against a configurable policy. The policy can define limits for maximum steps, runtime, estimated cost, and repeated actions. The core workflow is simple: text Agent proposes an action → Fuse evaluates the next invocation → Policy allows or blocks → Allowed action executes → Decision is persisted → Audit event is recorded
When an action would violate the policy, Fuse blocks the next invocation before the underlying executor runs. The run transitions to BREAKER_TRIPPED, the reason is recorded, and the complete decision remains visible through an ordered audit timeline. This distinction is central to Fuse. The system does not merely display a warning after an unsafe action has already occurred. It creates a pre-call control point that evaluates the proposed invocation before execution. The demonstration includes a synthetic invoice-verification loop. In this scenario, an agent repeatedly proposes the same vendor-verification action. Fuse detects the repeated normalized action and evaluates whether the next invocation would exceed the configured policy. Once the threshold is reached, Fuse produces: text BREAKER TRIPPED Next invocation blocked MAX_REPEATED_ACTION_EXCEEDED
The audit timeline records the sequence of events: the run begins, actions are evaluated, permitted actions execute, repetition is detected, the policy blocks the next invocation, and the breaker trips. This gives developers and operators evidence of not only what happened, but why it happened. Fuse also demonstrates the safe path. In the safe-completion scenario, the workflow remains within its configured boundaries, so Fuse allows the valid actions to execute and the run reaches COMPLETED. This shows that Fuse is not designed to block useful work. Its purpose is to stop unsafe continuation while allowing bounded workflows to finish normally.
The project is organized into separate layers to keep the safety mechanism understandable and testable. Shared contracts define the run, policy, event, and decision structures. The policy engine performs deterministic evaluation without relying on network calls. The enforcement layer controls whether a proposed action is allowed to reach its executor. The runner provides deterministic synthetic scenarios, while the persistence layer stores run state and ordered audit events. The dashboard exposes the result in a control-room interface. Fuse was designed with an AWS-native architecture using TypeScript, AWS Cloud Development Kit, and AWS SDK patterns. The planned serverless architecture uses Amazon API Gateway for HTTP access, AWS Lambda for API and runner execution, Amazon DynamoDB for run and audit-event persistence, Amazon EventBridge for breaker-trip event integration, Amazon CloudWatch Logs for operational logging, Amazon S3 and Amazon CloudFront for dashboard hosting, and Amazon Bedrock for future live model integration. The current demonstration is intentionally transparent about its status. It runs deterministic synthetic scenarios, and live Amazon Bedrock Converse API invocation remains disabled with: text
FUSE_LIVE_BEDROCK=false This is because AWS account verification and deployment access were not completed during the current submission process. Fuse does not present synthetic activity as live Bedrock execution, and estimated run cost is not represented as final AWS billing. The current MVP also has defined boundaries. Run history is session-local because a historical GET /runs endpoint has not been implemented. Policy contracts are read-only, with policy write endpoints outside the current scope. Fuse blocks the next integrated invocation; it does not cancel a call that has already started. These limitations are deliberate and documented. A safety product should be precise about what it guarantees. Fuse does not claim to eliminate every risk associated with autonomous AI. Instead, it provides one focused and verifiable control: evaluate the next action before execution, block it when policy requires, and preserve the evidence.
The broader vision is to make this control reusable across autonomous systems, including Bedrock-based agents, tool-using assistants, retrieval workflows, research agents, data-processing pipelines, and serverless orchestration systems. Autonomous AI needs more than intelligence. It needs boundaries, observability, and a reliable stopping mechanism. Fuse provides that starting point: 1)When the agent loops, Fuse evaluates. 2)When the policy is exceeded, Fuse blocks. 3)When the decision is made, Fuse records the evidence. 4)When the workflow is safe, Fuse lets it continue.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article