Browser UIs and thin agent gateways need temporary AWS credentials to call tool APIs — without shipping long-lived access keys in the SPA. Amazon Cognito Identity Pools (federated identities) exchange IdP tokens for STS creds scoped by IAM roles. Distinct from IAM Roles Anywhere (cert/on-prem machine identity), AppSync Cognito user-pool auth for GraphQL, and VPC Lattice service auth: this is federated identity → temporary AWS creds for human/UI/agent callers.

Run coding-agent fleets on AWS Batch: Fargate Spot compute, array jobs per PR, retry rules, timeouts and EventBridge Scheduler triggers.

Mid-turn coding-agent tools write lint logs, build caches, and hunk diffs that die with the turn — but classic S3 latency and request pricing make hot scratch feel sticky. Amazon S3 Express One Zone directory buckets give sub-millisecond first-byte for ephemeral tool artifacts in a single AZ. Distinct from S3 Object Lock (immutability) and EFS shared workspaces — this is latency/throughput for short-lived tool I/O.

Polling a coding-agent job every 500ms wastes API calls and still feels laggy. AWS AppSync GraphQL subscriptions push plan/tool/diff progress events to the UI over managed WebSockets with Cognito/IAM auth — without you operating an API Gateway WebSocket fleet. Distinct from API Gateway WebSockets and Lambda response streaming: AppSync is managed GraphQL + subscriptions for progress fan-out.

Flat RAG chunks lose the edges that matter for coding agents: who calls whom, which package pins which CVE, and how a symbol rename ripples. Amazon Neptune stores call graphs and dependency graphs so tools traverse vertices and edges with Gremlin or openCypher instead of hoping vector similarity reconstructs structure. Distinct from OpenSearch Serverless semantic scratch and GraphRAG bootcamp posts — this is managed graph DB + agent tool design for dependency reasoning.

Coding-agent accounts drift: public buckets for artifacts, wildcard IAM for "just one debug session," missing encryption on scratch queues. AWS Config Conformance Packs continuously evaluate packs of managed/custom rules and report noncompliance so you catch unsafe baselines before Security Hub becomes a ticket graveyard. Distinct from CloudFormation Hooks (block at deploy) and SCPs (org hard deny): Conformance Packs are continuous compliance telemetry + optional remediation across the agent account fleet.

Stream Bedrock tokens from Lambda: RESPONSE_STREAM function URLs, API Gateway REST streaming or InvokeWithResponseStream, with deadlines and auth.

Coding-agent tool microservices scattered across VPCs and accounts need authenticated service-to-service calls without a spaghetti of ALB rules and hand-rolled mTLS. Amazon VPC Lattice gives you a service network, auth policies, and connectivity so the planner can invoke the linter, RAG, and sandbox runners with IAM-aware allow/deny — not open security groups. Distinct from PrivateLink-to-Bedrock: Lattice is for your own tool services talking to each other.

Coding agents that emit CloudFormation or CDK will eventually propose a public S3 bucket, a wildcard IAM policy, or an unencrypted RDS instance. AWS CloudFormation Hooks evaluate those templates before resources are created or updated — so unsafe infra fails closed in the deploy path, not in a post-hoc Security Hub ticket. Pair Hooks with SCPs for org hard caps and GuardDuty for runtime detection.

You do not need a glue Lambda to fan DynamoDB Stream session events or SQS tool jobs into your coding-agent runners. Amazon EventBridge Pipes filters, enriches (optionally), and delivers straight to Step Functions, API Destinations, SQS, or ECS — so the stream-to-runner path stays declarative, fail-closed, and cheaper than another always-on mapper function. Distinct from our earlier Pipes→Lambda enrich pattern: here the target is the tool runner itself.