Author profile
Henry Bui
Engineering, CloudThinker
Henry Bui leads the engineering team at CloudThinker, building and continuously improving the agentic system end-to-end.
He writes about agent memory, runtime performance, generative UI, and incident response — the parts of the platform that make AI agents fast and useful in production.
Posts by Henry (10)
How To
How to Use CloudThinker for Incidents: Three Ways, One Night
Three ways CloudThinker handles incidents in one night: a known fix approved from a phone, an incident room for something new, and a chat question that never becomes a page. Each comes with the exact settings to configure it.
Product
Your On-Call Has Not Met Astra Yet
We held the Resolve harness fixed and swapped seven models across 21 live SREGym Kubernetes faults. GPT-6 Astra closed every gradable fault at $3.53 per resolution, and the post explains why a single leaderboard row, or token price, is the wrong way to pick an incident model.
Product
Introducing Resolve
Resolve joins signal intelligence and AI investigation with a Memory layer, so 13K raw events become 40 actionable clusters and investigation starts on its own. This post walks one 3 a.m. cluster from suppression to a confirmed root cause, with every remediation gated by Auto Mode.
How To
Eager Tool Calling: How We Cut Agent Latency by 50% on Long Tool Chains
Parallel tool calling overlaps tools with each other but still waits for the model to finish streaming. Eager tool calling fires each tool the moment its block is sealed, overlapping execution with generation: about 50% faster median task completion in production, with a reproducible open-source benchmark.
Product
Generative UI in Production: Lessons from a Pydantic-to-DSL Migration
How CloudThinker's agent dashboards went from 30-40 seconds and ~$0.50 per report to under 10 seconds and ~$0.08: why matplotlib and raw HTML were dead ends, why constrained-decoding JSON couldn't stream, and how a line-oriented DSL fixed both.
How To
Choose the Right Model, Optimize Your AI Costs
Running every task on one model tier either burns credits or fails the hard analyses. This post explains CloudThinker's Light, Pro and Ultra tiers, the break-even math behind building a skill with Ultra and running it on Light, and a 1,000-task routing example that saves 28 to 42% against all-Pro.
Product
Agent Memory Meets Graph: Introducing MemGraph, Long-Term Memory for AI Cloud Agents
A flat vector store can't tell which memories are related, outdated or proven. MemGraph stores operational knowledge as a typed graph, curates every new memory through an extract, search, evolve and connect pipeline, and ranks results with hybrid search plus reinforcement. This deep dive covers all five memory types, seven edge types and the four storage backends.
Product
Introducing CloudThinker Incidents
CloudThinker Incidents investigates alerts the way a senior on-call would: it gathers evidence, tests hypotheses, scores its confidence and maps the blast radius across your topology. You validate the conclusion and decide the fix.
Product
From Weeks to Hours: Building an AI-Powered Cloud Assessment Engine
A Well-Architected Review is sequential, a snapshot, and delivered as prose. This post explains how CloudThinker splits the review into independent pillar-by-resource conversations run in parallel, walks one right-sizing cell end to end, and shows why a weekly ten-minute assessment beats a yearly consulting report.
Product
CloudThinker Agentic Orchestration and Context Optimization
One agent with every tool and one growing history hits a wall on long-running cloud work. This engineering deep dive covers the supervisor pattern, delegation versus transfer, three-tier prompt caching, asynchronous self-summarization, consolidated tools and parallel tool calls, with the token math behind each.