https://kenkitts.com/Recent content on Ken KittsHugoenMon, 13 Jul 2026 00:00:00 +0000https://kenkitts.com/posts/bedrock-agentcore-part-5/Mon, 13 Jul 2026 00:00:00 +0000https://kenkitts.com/posts/bedrock-agentcore-part-5/The agent finally learns who it's talking to. Cognito JWT auth locks the gateway, an interceptor Lambda injects identity on a channel the model can't touch, and DynamoDB grows a composite key.https://kenkitts.com/posts/llm-gateway-part-2/Sat, 11 Jul 2026 00:00:00 +0000https://kenkitts.com/posts/llm-gateway-part-2/Five CDK stacks, a gateway that refused to talk to its own telemetry collector, an architecture mismatch that didn't surface until runtime, and an RDS version that didn't exist. The deployment worked. Here's what it cost.https://kenkitts.com/posts/llm-gateway-part-1/Sat, 04 Jul 2026 00:00:00 +0000https://kenkitts.com/posts/llm-gateway-part-1/Your developers are already using AI coding tools. Right now, on credentials you can't see, spending money you can't attribute. An LLM gateway is the corporate answer to that problem — but it comes with its own invoice. Here's how to decide if it's worth it.https://kenkitts.com/posts/bedrock-agentcore-part-4/Tue, 30 Jun 2026 00:00:00 +0000https://kenkitts.com/posts/bedrock-agentcore-part-4/For three posts the agent has saved your notes to a Python list that dies with the process. This post gives them a real home — DynamoDB behind a Lambda, served to the agent as MCP tools through an AgentCore Gateway. The agent can't tell the remote tools from the local ones. That's the trick, and the knife.https://kenkitts.com/posts/bedrock-agentcore-part-3/Sat, 27 Jun 2026 00:00:00 +0000https://kenkitts.com/posts/bedrock-agentcore-part-3/The agent runs in the cloud now, and it still forgets everything the instant the box recycles. This post wires it to AgentCore Memory — durable conversation that survives a cold start, then long-term memory so it remembers you across sessions. Two honest scars included, free of charge.https://kenkitts.com/posts/bedrock-agentcore-part-1/Thu, 25 Jun 2026 00:00:00 +0000https://kenkitts.com/posts/bedrock-agentcore-part-1/I already built an agent from scratch — a while loop and 600 lines of Python. It worked great, right up until I had to run it somewhere other than my laptop. This series is about that 'somewhere.' Part 1: what AgentCore actually is, and a Strands notes agent running locally in 40 lines.https://kenkitts.com/posts/bedrock-agentcore-part-2/Thu, 25 Jun 2026 00:00:00 +0000https://kenkitts.com/posts/bedrock-agentcore-part-2/In Part 1 the agent ran on my laptop, which is a great place for software to live right up until you close the lid. This post hands the exact same agent to AgentCore Runtime — one new file, no changes to the agent — and gets it running in the cloud behind a real endpoint.https://kenkitts.com/posts/agent-harness-part-5/Sat, 20 Jun 2026 00:00:00 +0000https://kenkitts.com/posts/agent-harness-part-5/A tool returns data. A skill returns instructions. Same dispatch loop, same tool_result block, completely different trust model. Let's build a progressive-disclosure skill system in 30 lines and then talk about why installing a skill is a form of prompt injection you chose.https://kenkitts.com/posts/agent-harness-part-3/Fri, 12 Jun 2026 00:00:00 +0000https://kenkitts.com/posts/agent-harness-part-3/The model forgets everything between API calls — so you carry the transcript, and you pay for it every single turn. Let's build the Session object, re-inject the plan to stop the model drifting, add a guard that won't let it fake 'done,' and stare at the token bill that memory quietly runs up.https://kenkitts.com/posts/agent-harness-part-4/Fri, 12 Jun 2026 00:00:00 +0000https://kenkitts.com/posts/agent-harness-part-4/Local tools are knives you forged yourself. MCP lets the model reach for knives someone else made, on someone else's server, behind an OAuth login. Let's wire a remote tool server into the same registry the local tools use — and stare hard at what it means to trust a tool you didn't write.https://kenkitts.com/posts/agent-harness-part-2/Wed, 10 Jun 2026 00:00:00 +0000https://kenkitts.com/posts/agent-harness-part-2/Tools are what turn a chatbot into an agent — and what turn a polite assistant into something with filesystem access. Let's build the decorator-based registry, wire it into Bedrock, and add the guardrails that stop the model from reading your SSH keys.https://kenkitts.com/posts/agent-harness-part-1/Mon, 08 Jun 2026 00:00:00 +0000https://kenkitts.com/posts/agent-harness-part-1/LLMs can only produce text. So how do they 'use tools'? Let's break down the agent harness — the code that turns a language model into something that can actually do stuff.https://kenkitts.com/about/Mon, 01 Jan 0001 00:00:00 +0000https://kenkitts.com/about/About Ken Kitts