Building Better Agents through End-to-End Evals

The biggest mistake teams make with agents is automating a process they don't understand. Lessons from building Ramp Stack's bookkeeping agent: map the real job first, then simulate business data so every stage can be graded end to end.

Memory Isn't Learning: Why AI Agents Keep Making the Same Mistakes

A senior analyst doesn't start from scratch every morning, but AI agents effectively do, and everything resets when a new teammate starts using them. Why storing memory isn't the same as learning from experience, and what it takes to build agents that actually improve.

Are You Smarter Than a Fruit Fly?

Google mapped the complete male fruit fly brain, and its 176,422-neuron simulation runs on a laptop CPU. So naturally, the next step was to make it do code review, with a TypeSafe model translating each change into something a fly can smell.

What a 48-Hour Hackathon Can Do for LLM Reliability

Getting an LLM feature working in a demo takes an afternoon; getting the same clean, structured answer every time is harder. Devpost's engineering team spent a 48-hour internal hackathon on the orchestration layer around the model rather than the model itself.

Beyond the Prompt Box: How Bloomberg is Architecting the Era of "Multiplayer AI"

Inside a 45-year-old technology estate with more than 9,000 engineers, producing code faster isn't enough. Bloomberg's head of architecture and emerging technologies explains how the company is moving past isolated prompt boxes toward shared, "multiplayer" AI.

A Data Company in an AI World

Moshi founder Anekha Sokhal traded energy derivatives at JP Morgan for nearly a decade before building a real-time quant platform in Houston. She talks about why she calls Moshi a data company, and the gap between how easy it is to build with AI and how hard it is to run what you've built.

An Agent's Plan Isn't Permission

An agent's credentials usually grant the same access no matter what it was asked to do. Intent-based authorization uses the purpose and context of each task, not the agent's plan, to decide what it's allowed to touch.

The Model Is the Easy Part

What does it actually take to move a coding workload off a frontier API and onto open-weight models you run yourself? Months of testing real software engineering tasks turned up three surprises, including that the agent harness changes the bill but barely moves resolve rates.

Your Financial Agent Passed Every Eval. Can You Defend Its Decision?

In audit, being right isn't enough; you have to show the basis for the opinion. Walking through an agent that tests bank reconciliations shows what defensible, agentic decision making has to look like.

For Agents, Waiting is the Hardest Part

In financial workflows, the danger often isn't the steps but the waiting between them, when deploys, crashes, and hardware failures pile up. AI agents add another source of delay, and durable execution is how to keep long-running work from falling apart.

Inside the $30-a-Week Self-Improving Agents Eliminating Enterprise Back-Office Friction

Enterprise AI usually assumes reliability costs money. Brainbase Labs' Milind Shah unpacks a Domino's case study where self-improving back-office agents delivered a 50-point accuracy jump while cutting costs from thousands of dollars a month to $20 to $30 a week.

The AI Data Paradox: The Safer You Make the Data, the Worse the AI Gets

In enterprise AI, performance is often decided before the model or prompt, by what goes into the pipeline. Sanitizing sensitive data strips out the very specifics that make it valuable, leaving AI systems working from a mottled blueprint.

Operational intelligence is sociotechnical

Deploying Clonal at a private equity firm involved far less talk of evals and models than of human factors and alignment, the human kind. Building software is the easy part; the hard part is systems people trust, use, and improve.

The "cool new stuff" trap

Shy Ruparel tried an AI agent to turn STL files back into editable CAD. What actually made the project work was much less glamorous: generate candidates, validate them, and keep the one that passes.

Nobody's Winning the AI Race Yet

Google just released Gemini 4 Argon, its first flagship since February, and reports of Gemini's death look greatly exaggerated. We're still early enough in AI that there's plenty of room for comebacks and setbacks.