Skip to content

Writing

Essays and technical notes

The essays are where I work out what I actually think. The technical notes are narrower and more practical: how to make a generative system produce something you would put in front of an enterprise customer, and how to prove it did.

Essays

August 2026

Essay

I Was Wrong About What Makes People Want an AI to Ask

Two weeks after arguing that agents are built to be correct rather than personal, I ran a small pilot — and the variable I had been designing around was not the one that moved people.

Almost nobody has one autonomy setting. Change rates looked similar across pairs; direction did not. File deletion produced the clearest shift toward oversight, cost did not, and stated rules did not always match what people chose. The useful object may be a function of person and situation, not a slider.

Read the essay →

August 2026

Essay

Reliable For Whom?

Ten years of watching what people actually do, and the question that broke my own work: I had spent years making agents reliable without ever asking who they were reliable for.

Preference data measures what an agent produces, and that converges — which is why you can average votes and rank models on a leaderboard. How the work should be done is a separate axis, and it diverges by person. Individual procedural preference is the one quadrant of the four nobody has measured, because it only became observable once agents began acting.

Read the essay →

Technical notes

May 2026

26 min

Measure Agent Performance With a Reproducible Evaluation Framework for MuleSoft Vibes

How do you know an autonomous agent got better rather than just got lucky? Agent behavior is not deterministic, so the usual software testing instincts quietly mislead you.

Three pillars: a benchmark of real enterprise integration work, a production-like IDE environment to run in, and an evaluation layer that reads execution traces rather than final answers. The unit of measurement becomes the distribution over repeated runs, not the run — and evaluating agents differs from evaluating software because what matters is how they reason, recover, and stay consistent.

Read on the MuleSoft blog →

September 2025

13 min

Trusted AI, Faster Outcomes: How MuleSoft Turns Generative Output Into Real Value

Building generative AI features has become routine, yet the overwhelming majority of pilots never produce measurable business value. Something between the demo and production is being skipped.

Validity and correctness are different problems, and correctness is the one that matters. A generated flow can be syntactically perfect and still miss the business logic completely. Grounding, validation, error correction and an LLM judge move roughly 20% and 17% to roughly 90% and 80%.

Read on the MuleSoft blog →

August 2025

17 min

DataWeave Generative Transformation Deep Dive

How do you generate correct code in a low-resource proprietary language that a general model has barely encountered, where the user often cannot describe the transformation precisely anyway?

Sample input and output pairs are a better specification than schemas, and they double as an executable oracle: compile the script, run it on the input, compare against the expected output. That turns error correction into a closed loop over a taxonomy of failure modes rather than a blind retry.

Read on the MuleSoft blog →

May 2025

12 min

Technical Guide to Einstein for Anypoint Code Builder: Generative Flows

How does a vague, multi-turn conversation become a working integration flow across an ecosystem of hundreds of connectors and thousands of operations?

Intent has to be consolidated before retrieval, not after. Summarize the whole conversation into one coherent prompt, then ground it by retrieving the relevant connectors, operations and examples — because retrieval quality is bounded by how well you understood the ask in the first place.

Read on the MuleSoft blog →

All technical writing on the MuleSoft blog