Bekah is a writer, community builder, and developer advocate. She founded Virtual Coffee, has worked across open source, developer education, and technical content, and brings a humanities-first lens to how people learn, build, and work with new technology.
bekahhw.com ↗Most teams manage AI cost by limiting it: caps, restrictions, chargebacks. Before you do that, measure how much frontier-model spend is routable to a cheaper model without changing the result.
Read more →A beginner's walkthrough of turning captured coding-agent sessions into a Fireworks fine-tuning run: pick one behavior, label it with paperctl, cut the sessions into examples, build a baseline, train, and judge the result with an LLM.
Read more →Picture an office building with no front desk: everyone has a key to every door, and nobody knows who's inside or what it costs. That's most companies' AI setup. An AI gateway is the front desk. The next two layers make the desk smarter: it starts asking what kind of work you're here for, and then it learns whether the work succeeded. The best way to use AI should also be the easiest.
Read more →Agentic development removed the friction that used to leak knowledge across a team. The work got faster, and keeping everyone caught up got harder.
Read more →Agent telemetry can price every session to the penny and still can't say what the work was for. Labels attach intent while it's still known, and a few months of them becomes a dataset of engineering work indexed by why.
Read more →The cost of an engineering decision is not the cost of making it. It's the work the decision causes afterward. We traced one architecture decision through 474 agent session records: $4,704.74 downstream of a $57 design doc, and four metrics (reach, duration, rework, amplification) any team with session records can track.
Read more →We analyzed two months of our team's sessions to find out what should become reusable knowledge. What we found is that repeatable work matters more than frequent work.
Read more →I audited our agent sessions expecting to find wasteful repeat calls. Repeats were only 4%. The real money was in agents re-reading what they already knew, plus a permission-check tax we could finally put a price on.
Read more →Your team keeps paying for the same solved problem, and because the tickets all close, nobody sees it happening. Repetition is evidence of a missing team capability, and standup is the detection layer that's been announcing it all along.
Read more →Before companies use AI cost overruns to justify cutting people, they should terminate the agent workflows that consume money without producing reliable outcomes.
Read more →One session, $100.66, 91% Fable 5 — for confirming rollout steps. Multiply that across 200 engineers and you have a budget conversation nobody is ready for. The paper console makes the numbers visible.
Read more →Evidence gets a skill written. Evaluation keeps it worth keeping. A skill that isn't measured can silently make your agent worse, and, because skills execute, a wrong one does more damage than a wrong doc ever could. Here's why every skill needs an eval, and what a real one looks like.
Read more →Skills are about to become part of the AI engineering stack. If they don't carry provenance, we're just rebuilding prompt libraries with nicer names. A prompt starts with intention. A skill should start with evidence.
Read more →Frontier labs are paying a billion dollars a year for the same kind of session reasoning your team generates and deletes every night. The asset isn't the code the agent produced. It's the path the agent took.
Read more →AI agent workflows are no longer single requests. They are stateful, branching systems that mutate their own context over time. Request logs, traces, transcripts, and bills flatten that structure. Teams need telemetry primitives that preserve session lineage, cache topology, and prompt shape.
Read more →