Posts
All posts
Agents on Call, Part 8. Production: Observability, Evals, and the Day It Lies
OTEL traces into CloudWatch, Bedrock invocation logging to S3, an eval harness with a golden incident set, and the day the triage agent lied with confidence.
Kiro After the Hype: What AI IDEs Actually Changed
Eight months after Kiro went GA, the spec survived and the IDE mostly did not. What actually changed is where the review happens, not who writes the code.
Why Your AI Pilot Died in Procurement
Most AI pilots do not fail on accuracy. They stall in security review and data processing agreements, because nobody scheduled the twelve weeks that follow.
Per-App Bedrock Cost Tracking with Inference Profiles
Application inference profiles put cost allocation tags on Bedrock calls, turning one shared bill into per-team lines. The tag design is the hard part.
Agents on Call, Part 7. Sizing: Token Math Nobody Does Upfront
The sizing math nobody does upfront: tokens per incident, quota ceilings, when provisioned throughput breaks even, and the platform's monthly bill.
EU AI Act, August 2: The Deadline That Didn't Move
The Digital Omnibus deferred the high-risk deadlines to 2027 and 2028. Article 50 transparency still lands on 2 August 2026, and that is the one most teams hit.
SageMaker vs Bedrock: An Org Decision, Not a Technical One
The SageMaker or Bedrock question is really about whether your org has a team that owns models. Pick for the team topology you have, not the one on the slide.
Prompt Injection via Your Own Docs: The RAG Attack Surface
Your knowledge base is untrusted input. Retrieval hands attacker-authored text to the model, so the control that pays off is scanning at ingestion time.
M365 Security 101: AI Pilot and Business Impact Reports
Where AI earns its place in security: remediation behind a per-change approval gate, and reports leadership can act on. A 101 with Aether365 as the example.
Agents on Call, Part 6. Guardrails: The Part Everyone Skips
Threat model, then guardrails: Bedrock Guardrails in Terraform, why IAM still does the heavy lifting, and the failure modes nobody demos on stage.
Trust the Model, Audit the Binary
A coding agent's client is the most privileged binary on your machine. Claude Code's hidden prompt fingerprint shows why you audit it, not trust it.
AWS Monthly (June '26): Agents Get a Feedback Loop
June 2026 on AWS: the Summit in New York turned production traces into agent improvement, shipped Continuum for security, and took the AgentCore harness to GA.
Agents on Call, Part 5. The Team: Supervisor and Three Specialists
A supervisor delegates to triage, runbook, and cost agents over the network, AgentCore Memory ties their findings together, and when one agent still wins.
AWS Built a Sandbox for AI-Generated Code: Lambda MicroVMs
AWS Lambda MicroVMs give AI agents a place to run model-generated code, isolated at the VM level. The missing piece for production agents was the runtime.
Your Multi-Agent System's Real Limit Is Tokens Per Minute
Amazon Bedrock now exposes per-model tokens-per-minute quotas in Service Quotas. For agents, TPM is the real scaling ceiling. Plan for it before the 429s start.
Estonia Is Giving AI Agents an ID. That Is the Easy Part
Estonia plans to issue AI ID codes to AI agents, a world first. The identity is the easy part: authority, delegation, and accountability are the real work.
Le Chaton Fat: The Fattest AI Model That Never Existed
There is no AI model called Le Chaton Fat. No weights, no API, no benchmark. Here is how a Mistral cat joke became the AI community's favourite running gag.
Agents on Call, Part 4. Tools and the Gateway: MCP, Allowlists, Read-Only Default
Tools move behind AgentCore Gateway: scoped Lambdas, cross-account read-only roles, and the one gated path that is allowed to change anything.
AI Coding Agents Need Staging Environments Too
Teams gave coding agents production credentials and called it velocity. An agent is a contributor with no judgment, and it needs the same environment ladder.
Agents on Call, Part 3. First Agent: Incident Triage in Strands
The first agent ships: a Strands triage agent on AgentCore Runtime, its evidence-first system prompt, two read-only tools, and the Terraform that deploys it.
Claude Code in CI: Letting an Agent Fix the Build
Running a coding agent in your pipeline is easy. The engineering is in the guardrails: what it can touch, when it stops, and who reviews what it proposes.
Agents on Call, Part 2. The Foundation: Terraform Before Tokens
Before a single Bedrock token flows: the ops-tooling account, the two spoke IAM roles, model access, and the inference-profile decision, all in Terraform.
The Context Window Is Not Your Friend
A huge context window is not a replacement for retrieval. Recall degrades as the prompt grows, cost scales with every token, and the middle gets skimmed.
AWS Monthly (May '26): Agents Get a Wallet
May 2026 on AWS: AgentCore Payments lets agents transact, and the Agent Toolkit for AWS plus a GA managed MCP server harden the toolchain that builds them.
Agents on Call, Part 1. The Scenario: Why an Ops Team Hires Agents
A mid-size SaaS drowns in on-call toil and decides to hire agents. Part 1: the scenario, the stack decision, and the AWS Bedrock architecture we will build.
Logging Prompts Without Logging PII
You need prompt logs to debug an LLM app but cannot keep raw PII in them. Redact before storage with Comprehend, then set retention so logs age out on schedule.
When Haiku Beats Opus: Model Right-Sizing on Bedrock
Defaulting every call to Opus is how LLM bills balloon. Route by task class: Haiku for the mechanical majority, Opus as the escalation path for hard cases.
Agentic RAG Is Mostly Latency You Don't Need
Agentic RAG loops through retrieval hops, each a model round trip. For most questions one good query wins. Reach for the loop only when it earns the latency.
Evals Before Agents: You Can't Ship What You Can't Score
Without an eval harness, every agent change is a vibe check. Build the scoreboard before the agent, and treat the LLM-as-judge as a component that can be wrong.
Semantic Caching: Two Different Questions, One Answer
Semantic caching returns one stored answer for two differently worded questions. It cuts cost and latency, but a false hit serves a confidently wrong reply.
Step Functions Is the Most Underrated Agent Orchestrator
Most agent workflows do not need a model to decide control flow. Step Functions gives deterministic orchestration, retries, and human-in-the-loop states.
LLM Gateways: Why Every Platform Team Builds One Eventually
A second team calling a model means ungoverned fan-out. An LLM gateway centralizes auth, quota, routing, and audit. Build vs LiteLLM vs API Gateway.
AWS Monthly (Apr '26): OpenAI Lands on Bedrock
April 2026 on AWS: OpenAI models, Codex, and Managed Agents arrive on Bedrock, and AgentCore adds a managed harness and CLI to shorten the path to an agent.
Cross-Region Inference: Cheap Resilience or Residency Trap?
Bedrock cross-region inference smooths throughput and throttling. But a global profile can route your prompt out of its geography. Read residency first.
Your LLM Bill Is an Observability Problem
A surprising Bedrock bill is not a pricing problem, it is a visibility one. If you cannot attribute tokens to a feature, tenant, or agent, you cannot manage it.
Batch Inference on Bedrock: Half Price If You Can Wait
Amazon Bedrock batch inference runs at 50 percent of on-demand pricing. The only cost is latency. For any job where nobody is waiting, that trade is free money.
Multi-Tenant LLM Apps: Isolating Customers on a Shared Model
One shared Bedrock model, many customers. The model is stateless, so isolation is your job: scope the retrieval, cap the quota, carry identity per tenant.
Agent Memory Is a Database Problem, Not a Prompt Problem
Stuffing an agent's whole history into the prompt is not memory, it is a growing bill and a token ceiling. Real memory is a database with a retrieval step.
Structured Output Beats Clever Parsing
Still regex-parsing JSON out of model text? Stop. Bedrock structured outputs enforce a JSON Schema during decoding, so the response is valid by construction.
Prompt Caching on Bedrock: The 90% Discount Most Teams Ignore
Bedrock prompt caching reads a repeated prefix at 90 percent off, but a cache write costs more than not caching. The breakpoint decides which you get.
AWS Monthly (Mar '26): Governance Comes for the Agents
March 2026 on AWS: AgentCore Policy and Evaluations reach GA, Elemental Inference ships, and agent governance moves from demo to a production control plane.
Streaming Responses Are a UX Decision, Not a Performance One
Streaming model responses is a user-experience choice about time to first token, not a speed fix. Sometimes it makes structured output and tool use worse.
Bedrock Agents vs Rolling Your Own Loop
Amazon Bedrock Agents handle orchestration, memory, and tool calls for you. Here is when the managed framework saves you real work and when it quietly owns you.
IAM for LLM Apps: Least Privilege When the Caller Is a Model
When the caller is a model, least privilege still applies. Give each agent tool a scoped IAM role and a session policy, not one broad set of admin credentials.
Someone Registered antrophic.com and Points It Straight to OpenAI
Someone registered antrophic.com, one letter off the real domain, and pointed it straight at OpenAI. A look at the bait and at AI brand confusion.
Stop Fine-Tuning. You Need RAG, a Cache, and Better Prompts
Fine-tuning plus provisioned throughput is the expensive answer to most LLM problems. The cheaper path is retrieval, prompt caching, and better prompts.
Knowledge Base Chunking Is Where Your RAG Quality Dies
Most bad RAG answers are a retrieval problem, not a model problem. How fixed, semantic, and hierarchical chunking in Bedrock Knowledge Bases set your quality.
Bedrock Guardrails Won't Save You From Prompt Injection
Amazon Bedrock Guardrails filter content, they do not authorize actions. Real prompt injection defense is input isolation, tool allowlists, and IAM scoping.
Cutting Amazon Bedrock Knowledge Base Costs by ~90%: Migrating from OpenSearch Serverless to Aurora Serverless v2 with pgvector
An OpenSearch Serverless vector store costs roughly $700/month before you ingest a document. Aurora Serverless v2 with pgvector drops that floor under $50.
AWS Monthly (Dec '25): The Kiro Era Begins
We ended the year with the General Availability of Kiro (Frontier Agents). Kiro is not just a chatbot; it’s a Virtual Software Development...
AWS re:Invent 2025: The "Agentic" Era
re:Invent 2025 read through an agentic lens: the Nova 2 family split into specialised roles, and AWS stopped shipping chat and started shipping work.
AWS Monthly (Oct '25): Industrializing AI Training
October 2025 on AWS: Project Rainier put over 500,000 Trainium2 chips behind one cluster, and Trainium2 price-performance now beats comparable GPU instances.
AWS Monthly (June '25): S3 Becomes Your Vector DB
For the last two years, we have been told we need a specialized vector database (Pinecone, Milvus, etc.) for Retrieval-Augmented Gene...
AWS Monthly (Feb '25): Automated Code Evolution
February 2025 on AWS: Q Developer moved from code completion to autonomous refactoring, mapping repository dependencies and migrating legacy Java itself.