Skip to content
SECTION — JOURNAL 102 ENTRIES

The full archive

The journal

Building in the AI-native era — where cloud meets coding assistants and agentic AI.

SERIES — ONGOING

See all →
AI & LLM Fundamentals Cloud Security AWS Architecture Agentic AI

SECTIONS — BY DOMAIN

All AIAI EngineeringCloudSecurityWeekly Digest

SECTION — AI ENGINEERING · 33 ENTRIES

AI ENGINEERING

27 October 2026 · 11 MIN READ

30 Failed Attempts and $21.87: The Output Cap of My Agent Factory Was Counting the Model's Thinking

At its identity step my agent build factory asks the model for three homepage mockups per site. For one site, five launches in a day produced 30 failed attempts, ten minutes of waiting each and $21.87 of model spend for nothing, while another site got its mockups the same afternoon. The cause was one number set months ago: a 32,000-token cap on each answer, on a model that now reasons before it writes, and whose reasoning counts against that cap. Here is how I found it, the per-model cap that fixed it, the replay that proved it before the merge, and the second defect the same morning exposed: an e-mail the page promised and nothing sent.

Read →

AI ENGINEERING

26 September 2026 · 11 MIN READ

AI-DLC in a Team: the Agent Is No Longer the Bottleneck

In January I used AI-DLC alone, one developer and one agent in an IDE. This month I deployed AIDLC Collaborative, the open-source sample that runs the same method for a whole team: intents, human gates, parallel construction lanes, agents in isolated AgentCore sessions, and a platform that owns git. Two intents with the same 48 minutes of agent work took 81 and 284 minutes end to end. Here is what the team version changes, what it costs to stand up, the failure that painted 14 empty stages green, and what I have not verified yet.

Read →

AI ENGINEERING

26 September 2026 · 11 MIN READ

Kiro Trusts One File. Here Is How I Stopped Editing It by Hand

Kiro's enterprise MCP registry is one JSON file behind one URL: whatever is in it, developers see; whatever is not, never starts. That makes the file a control point — and a hand-edited control point is a liability. I rebuilt mine as a generated artifact: AWS Agent Registry as the approved catalog, AgentCore Gateway as the only remote MCP endpoint, and a small publisher that turns approvals into the file Kiro reads. Here are the five decisions, the six things the docs did not tell me, and what I still have not verified.

Read →

AI ENGINEERING

22 September 2026 · 9 MIN READ

One 'AI Gateway', Two Different Jobs: the LLM Gateway and the Tool Gateway

Every agentic architecture conversation I have eventually collides with the phrase 'AI gateway' — and half the time the two people using it are talking about two different products. One proxies model traffic; the other fronts tools. Confusing them produces bad verdicts, broken migrations, and gateways justified by the wrong criterion. Here is how I untangle the two roles, the per-model API compatibility trap, and the one question that decides whether an extra gateway earns its place.

Read →

AI ENGINEERING

16 September 2026 · 10 MIN READ

Deck-as-Code: How a Fleet of Review Agents Saved My Conference Deck

Iterating a 24-slide conference deck under review pressure, I stopped editing PowerPoint by hand. Every version is now rebuilt by an idempotent python-pptx script from the last validated file, and reviewed by four parallel critique agents reading rendered frames — because the XML lies about what the eye sees. Eight blocking defects on a version I was ready to ship, one vertical-tab pitfall, and a ghost footer hiding in the slide master: here is the pipeline.

Read →

AI ENGINEERING

16 September 2026 · 9 MIN READ

Five Agents, One Football Team: Per-Role Model Selection and Deterministic Guardrails on AgentCore

I entered an agent football tournament: five LLM-driven players on Bedrock AgentCore, one command per game tick, every tick latency-sensitive. Per-role Nova model selection, memory switched off, and deterministic rules applied after the LLM made the team work — and then command fingerprinting in CloudWatch proved my perfectly healthy runtimes had been playing for the wrong team all along.

Read →