Fetching from the wire…
Public story · 2026-09-02 · high
A Lambda job rewrites IAM policy every 15 minutes so a blown budget blocks the next model call without anyone re-authenticating.
Why now: AWS published the architecture on its machine learning blog on September 2, 2026.
Jamf cuts an engineer's access to Claude models within minutes of a blown spend cap, without forcing a fresh login, per AWS. Hundreds of engineers stay covered for under $10 a month, built from AWS services the company already pays for.
AWS lays out the mechanics in its Bedrock writeup. Invocation logs stream to S3 carrying model ID, token counts, and the calling user's IAM Identity Center identity. An Athena view turns that into per-user daily spend. Every 15 minutes, a Lambda function on an EventBridge schedule checks who's over budget and rewrites their Customer Managed Policy with iam:CreatePolicyVersion. IAM evaluates the new policy on the user's next API call, so the block takes effect immediately. No token revocation, no session kill, no support ticket.
The system doesn't cut a user off outright. It denies Opus once someone crosses 80% of their daily budget, denies Sonnet at 100%, and leaves a cheaper model reachable the whole time. An engineer who burns through budget on the expensive model keeps working, just on something that costs less.
AWS's post doesn't say what happens at larger log volume, or whether the 15-minute window holds if usage spikes faster than that.
Each link below shares sources, entities, or timing with this story.
Two concrete recipes for regulated customers who need inference in a single region, not merely in-geography, since cross-Region inference is the throughput-friendly default. Path one: CLAUDE_CODE_USE_MANTLE=1 plus AWS_REGION, pinning models by plain ID, supported in Ireland, S...
AWS shipped per-user, per-target rate limiting August 6 covering MCP targets, inference targets and HTTP passthrough across three metrics: requests (RPS/RPM), tokens (TPM, inference only), connections (CPS). Limits scope by JWT claims ($.context.jwt.sub, .role, .azp) or IAM id...
Generally available August 31 in five regions, registering MCP servers with their tools and prompts, A2A agent card definitions, agent skills written as markdown with associated code, and custom JSON descriptors (AWS). It exposes its own search through an MCP server so agents...
CloudWatch subscription filters match ERROR/Exception/FATAL/CRITICAL, trigger a Lambda, hand off to a Strands Agents SDK agent running Claude Sonnet that pulls surrounding log context and the relevant GitHub source, then publishes a structured RCA to SNS for email and Slack (A...
The full family, Sol, Terra, and Luna, is generally available on Amazon Bedrock with IAM and VPC controls (LLM Boss, AWS). Sol targets coding, biology, and cybersecurity agentic work. Terra runs everyday tasks at about half GPT-5.5's cost, and Luna optimizes for speed. The thr...
Loom (github.com/awslabs/loom) is a reference implementation: "paved path" blueprints baking in least-privilege IAM, abstracted configuration, and strict deployment guardrails for agents built on Strands Agents running on Bedrock AgentCore Runtime. The repo ships the specs use...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.