Blog

Field notes & essays on quiet automation.

Short reads on building AI agents, choosing the right tool for the work, and the things teams get wrong before they call us.

StrategyJuly 20, 20265 min read

NeMo Automodel + Diffusers: When Vision Fine-Tuning Is a Trap

NVIDIA NeMo Automodel makes Diffusers fine-tuning easy. For agent teams processing video and images, that's exactly the problem. Here's when to skip it.

Read

EngineeringJuly 17, 20267 min read

Cache-Reason Logs: The Observability Pattern Your Agents Need

Vercel logs *why* a cache missed, not just that it did. Here's why agent tool calls and retrievals need the same granularity — or you're flying blind.

Read

StrategyJuly 15, 20266 min read

Useful Work Per Dollar Is the Wrong First Question for Agents

OpenAI's 'useful work per dollar' skips the harder question: does your problem need an agent loop at all? Here's how we measure the real tradeoff.

Read

EngineeringJuly 13, 20264 min read

Seedream 5.0 Pro Can't Ship Your Infographics Alone

Seedream 5.0 Pro's text rendering is a leap forward, but production agents still need SVG+image hybrid pipelines. Here's how we architect them.

Read

StrategyJuly 10, 20265 min read

GPT-5.6 in Copilot: Why Cost-Per-Token Is the Wrong Metric

GPT-5.6 in M365 Copilot changes agent economics. Why cost-per-successful-action beats cost-per-token — and what to fix in your retry logic first.

Read

StrategyJuly 8, 20266 min read

Eve Chat SDK Adapters: Easy Deploy, Hard Decisions

Eve's Chat SDK makes multi-surface agent deploys plug-and-play. Here's what teams miss: per-platform rate limits, state divergence, and when a webhook wins.

Read

EngineeringJuly 6, 20267 min read

HuggingFace Kernels: Why Agent Teams Should Care About the Iteration Loop

HF's revamped Kernels aren't just about custom hardware. For agent teams, the real win is a faster loop from inference bottleneck to shipped optimization.

Read

EngineeringJuly 3, 20266 min read

Vercel AI Gateway Routing Rules: Necessary, Not Sufficient

Vercel's AI Gateway routing rules move failover out of code deploys. Here's what they actually solve for agents at scale—and what they don't.

Read

EngineeringJuly 1, 20266 min read

ScarfBench Is the Canary: Why Java Migrations Predict Agent Reliability

ScarfBench doesn't rank models—it exposes where agents break on real codebases. Here's why enterprise Java migration predicts your production agent's reliability.

Read

EngineeringJune 29, 20264 min read

HP + OpenAI: The Real Bet Is Agent Observability, Not the Model

HP's OpenAI Frontier partnership isn't a model story. It's an operations story. Here's the agent observability and routing infrastructure enterprises actually need.

Read

StrategyJune 26, 20265 min read

HF Jobs Made vLLM Easy. That's Not the Decision You're Making.

HuggingFace's one-command vLLM deploy is easy. Here's the cost, latency, and lock-in math that decides whether you should actually self-host inference.

Read

StrategyJune 24, 20265 min read

GPT-5 Pro Cracked an Immunology Mystery. Should You Copy the Pattern?

GPT-5 Pro solved a 3-year immunology puzzle. Here's the engineering pattern that made it work — and when to skip frontier reasoning for a cheaper workflow.

Read

StrategyJune 12, 20264 min read

At 4B Tokens/Day, Your Agent Stack Is a Routing Problem

Okara pushes 4B daily tokens across multiple providers on Vercel. The lesson for agent teams: at scale, orchestration is solved—token routing isn't.

Read

StrategyJune 10, 20266 min read

Nemotron 3.5 Safety: Why One Filter Won't Hold Your Agents

NVIDIA Nemotron 3.5 Content Safety is customizable—and that's the trap. Why agent safety has to be per-decision, and what it costs to do right.

Read

StrategyJune 10, 20266 min read

Why "Just Add an Agent" Killed Amazing Digital Dentures

Amazing Digital Dentures bolted an LLM agent onto a supply chain it didn't own. Here's why agents fail without deep domain integration—and what to build instead.

Read

EngineeringJune 10, 20266 min read

Code-Switched ASR Breaks Voice Agents. Here's the Fix.

Frontier ASR still fails 2-3x harder on bilingual speech. The fix isn't waiting for better models — it's runtime code-switch detection and language routing.

Read

AgentsJune 1, 20265 min read

NVIDIA Cosmos and the New Vision Tax on Physical-World Agents

Cosmos makes video-based world models cheap to access. The real question for your agent stack: when is the vision tokenization overhead worth paying?

Read

EngineeringMay 29, 20267 min read

What Endava's 'Weeks to Hours' Actually Required Under the Hood

Endava cut requirements analysis from weeks to hours with Codex. The real story isn't agents—it's the workflow engineering most teams skip.

Read

EngineeringMay 25, 20265 min read

Diffusion LLMs Are Fast. Your Agent Is Still Slow.

Nemotron-style diffusion LLMs cut decoding time, but agent latency lives in retrieval and tool calls. Here's what actually changes in production.

Read

EngineeringMay 21, 20266 min read

Your RAG Hits 0.9 Recall in Eval and Dies in Production. Here's Why.

Benchmark recall lies. Here's how we diagnose retrieval quality in production RAG systems before users start filing tickets at 2am.

Read

AgentsMay 21, 20267 min read

Stop Defaulting to GPT-4o: Right-Sizing Your Agent's Reasoning Loop

OlmoEarth v1.1's efficiency gains expose a habit: overprovisioning frontier models. Here's how to audit token spend and route only the hard cases upstream.

Read

StrategyOctober 14, 20256 min read

When to build an agent vs. a workflow

Agents and workflows look similar from the outside. They cost very different amounts to maintain. Here's how we pick.

Read

StrategyAugust 21, 20255 min read

The hidden cost of manual ops

Manual operations rarely show up in a budget. They show up in the people who quietly leave.

Read