The Model Was the Same. The Serving Configuration Was Not.
I ran official Qwen3.8 27B BF16 and FP8 side by side, then changed one serving knob at a time. The fastest result was not the whole story.
Sekou M. Doumbouya - Technical Insights
Technical lead and Senior Staff Software Engineer with 20+ years building cloud infrastructure, network architecture, reliability systems, and agentic developer tooling.
Latest posts
What Wins in a Benchmark Is Not What You Ship
The best long-context profile I measured passed every synthetic gate and crashed twice under real traffic. Here is why promotion has to be a guarded transaction, not a config edit.
A Card That Made Me Sign a Letter
Why I added a second RTX PRO 6000, upgraded my CPU, and then watched DeepSeek 0731 land on the exact card I just bought.
The Task Is Not the Unit of Work
A 41-task graph ran for 41 hours, produced 115 reviews, and merged nothing. Then I pivoted mid-flight to milestones, and the same session shipped five PRs in half the time.
Structurally Valid Is Not Good Enough: Why My Model Router Measures Quality
anvil-serving routes coding-agent work to local models only where measured quality has earned it, verifies local output, and escalates up the tier ladder the moment verification fails.
The Model Name Is Just a String
How wiring two mismatched GPUs into my coding harnesses turned into intent-based routing, and why the measurement, not the transport, became the product.
A Billion Tokens to Ship One Feature. Forty-Six Million Were Real.
Two days of autonomous agents shipped a release and moved a billion tokens. Only about five percent was generated work. That five percent is the number that sizes a local box.
I Bet on Owning My Own Substrate
Why I built a dual-GPU workstation and stood up local model serving, and why owning the durable layer is the bet I am most glad I made.
I Said 'Report in the Morning' and Went to Sleep. It Shipped a Release.
I handed an agent a goal and one sentence, then slept. 23.7 hours of unattended runtime later, Anvil v0.3.0 was on PyPI. Here is what it woke me for, and what it shouldn't have.
Status Is Downstream of Proof: The Architecture of Anvil
An agent reported a test it never wrote. That gap is the whole reason Anvil exists, now open source as anvil-state on PyPI. Here is the architecture, and the one part still unsolved.
I Rebuilt My Crew on a New Runtime. Only the Ledger Survived.
I ported a nine-agent crew from one harness to another. The orchestration had to be rebuilt. The ledger moved untouched. That gap is the whole product.
I Didn't Need a Bigger Computer. I Needed Room to Think.
A single 5090 could run three local agents. It just couldn't give them room to work. The story of the VRAM wall, the three ways out I priced, and the upgrade I actually made.
Volume Goes Local. Stakes Go to the Cloud.
The question that routes an agent task is not which model is best. It is what the task can afford to get wrong. How I pinned a nine-agent crew across a cloud frontier tier and a local GPU.
I Bought a GPU to Run My Own Agents. Then I Red-Teamed the Plan.
Maxing my AI plans and fighting rate limits, I split the spend across providers, bought a local GPU for overflow, and designed a pull queue for my agents. Then I red-teamed the plan.
Everyone Grew the Same Eye: A Field Guide to the Spec-Driven Wave
Spec Kit, Kiro, Tessl, task-master, BMAD, Agent OS, OpenSpec: seven projects, twelve months, one idea. A field guide to who's building what, and why they converged. Part 2.
The Newest Idea in AI Coding Is Forty Years Old
Spec-driven development feels like a 2025 invention. It isn't: it's literate programming, MDA, TLA+, BDD, and Terraform's desired state in a hoodie. Part 1 of a series.
Where the Wave Breaks: The Layer Nobody Wants to Build
Every spec-driven tool stops at the same place: a great document handed to an agent, and trust. The layer they skip (durable state, leases, evidence) is the hard part. Part 3.
The Anvil, Not the Hammers: Where fakoli-state Goes Next
I had an agent system audit my agent system against the field. The verdict: the orchestration is being commoditized. The moat is the boring, durable state layer.
Reading Other People's Bug Trackers: The Bets Behind fakoli Are Getting Called
I mined four hundred GitHub issues from the tools next to mine. The field's pain validated the bets — and Anthropic shipping agent teams natively confirmed the biggest one.
The Fakoli Style: An Operating Model for Building With Agents
Three plugins, four invariants, one principle. What I built isn't a toolkit — it's an operating model for serious development with AI coding agents.
Plan, Claim, Apply: Building fakoli-state
A local-first project-state layer in SQLite, with claims that have heartbeats and evidence captured by hooks. Here's how it was built and which trade-offs I'm still living with.
State Is the Product: What Agentic Workflows Were Missing
Workflow orchestration alone won't save you. The thing that breaks first when you run real multi-agent work is the state — and almost nobody is building that layer.
Building the Engine: Architecture of a Durable Agentic Execution System
A 10-package monorepo, 27 MCP tools, checkpoint-based durability, and three sandbox types. The full blueprint for Baara Next.
Intent Should Be Durable: The Missing Layer Between Agent SDKs and Production Systems
Agent SDKs give you tools and reasoning. They don't give you queues, checkpoints, or crash recovery. That gap is where your intent dies.
From Pressing Buttons to Intent-Driven Flow
How I went from managing AI agent context windows by hand to building an orchestration system that lets me describe what I want and walk away.
Teaching AI Agents to Work Like a Team (Not a Crowd)
What I learned building specialized agent archetypes that coordinate through file ownership, wave execution, and structured handoffs.
From Cloud Infrastructure to AI Infrastructure: What Transfers and What Doesn't
After twenty years building cloud platforms, here's what I've learned about bridging into AI infrastructure, and which skills matter more than you'd expect.
Why I Rebuilt My AI Assistant From Scratch (And What the Analysis Told Me)
How a structured competitive analysis of my own code convinced me that a clean-room rewrite was the right call, and what I learned building Nexus.
Documentation Is Infrastructure (Just Made of Text Instead of Routers)
Why the blueprint document I wrote for our AWS organization might be the highest-leverage thing I shipped that year.
The Multiplier Effect: Why I Keep Mentoring as a Core Output
The only sustainable scaling strategy is multiplying competent decision-makers. Here's how I think about mentorship as a senior IC.
Platform Engineering Is Product Management for Infrastructure
Why treating your internal platform like a product changes how infrastructure teams operate, and how golden-path solutions reduce cognitive load.
Questioning Architectural Defaults: A Transit Gateway Cost Story
Why periodically questioning inherited infrastructure decisions matters, illustrated through an AWS Transit Gateway to VPC peering migration.
The Player-Coach Model: Why I Stay 40% Hands-On
Why staying hands-on at the principal engineer level creates better outcomes for your team, your architecture, and your credibility.
Designing Multi-Region Infrastructure: Patterns That Scale
A taxonomy of region types, spec-driven provisioning, and phased orchestration for predictable multi-region expansion.