Goldman's CIO Stops Counting Seats and Starts Counting Cycle Time
Marco Argenti, Goldman Sachs' tech chief, has decided that per-engineer AI usage metrics tell him almost nothing useful. With 12,000 engineers working alongside an in-house ChatGPT-style assistant, the question that matters is how quickly an idea becomes a production deployment. Tool adoption dashboards, in his telling, are vanity metrics dressed up as governance.
This is a meaningful reframing for engineering leaders trying to justify Copilot or Cursor budgets to finance. Counting prompts or accepted suggestions rewards activity, not throughput. Cycle time from ticket to merged-and-deployed code captures whether the tools are actually compressing the work, including the parts AI is bad at: review, integration, and rollback.
The practical takeaway for individual engineers is that the next wave of internal evaluation will likely look less like 'did you use the assistant' and more like 'did your PRs ship faster and stay shipped.' That favors workflows where AI is embedded in the boring middle of the SDLC, not just the green-field code generation moment.
A Small Library Pointing at a Big Agent Problem: Local Auth
Two releases of authsome landed on PyPI in quick succession, moving from 0.2.3rc67 to 0.2.4. The project bills itself as a portable local authentication library for AI agents and developer tools. That description is doing a lot of work, because authentication for autonomous agents is one of the genuinely unsolved problems in the current stack.
Most agent frameworks today punt on credentials. They either run with the developer's full session token, or they wire in a service account with privileges far broader than any single task requires. Neither approach scales to a world where agents fan out across APIs, file systems, and internal services on behalf of different users. A portable, local-first auth primitive sidesteps the cloud-broker model that has dominated agent identity discussions so far.
It is worth watching libraries at this stage of maturity precisely because they are not yet opinionated. The release cadence (an rc to a point release within days) suggests the surface area is still settling. Engineers building agent tooling should pay attention to how this category evolves, since whichever pattern wins will shape how agent permissions get audited for the next decade.
Answer Engines Are Quietly Rewriting the Developer Tools Market
The shift from search to answer engines is usually framed as a marketing problem, but its second-order effects land squarely on developer tooling. When buyers no longer click through to documentation and instead receive a synthesized response from an LLM, the surface that wins mindshare is whatever the model has been trained on or can retrieve cleanly.
That changes the calculus for any team shipping a developer product. Documentation is no longer a static artifact for humans to skim; it is training data and retrieval fodder. Pages that read well to a model (clear headings, canonical examples, unambiguous terminology) increasingly outrank pages that read well to a developer skimming on a Tuesday afternoon. SDK authors are starting to write with both audiences in mind.
The reverse pressure is also real: marketing tools themselves are being reshaped to handle brand interpretation and data synthesis rather than keyword campaigns. For engineers, the lesson is that the same retrieval and grounding patterns being deployed to influence answer engines on the marketing side are the patterns worth studying for internal RAG systems. The tradecraft is converging.
A 25% Retention Lift from Widgets, and What It Says About AI Surfaces
Gratitude, a journaling app, reported a 25% retention bump after investing in Android home-screen widgets. The number is striking on its own, but the more interesting question for AI engineers is what it implies about ambient surfaces. Widgets succeed because they remove the friction of opening an app; the user is reminded and engaged in the same gesture.
Agent-driven products face the same friction problem, amplified. A chatbot buried behind a launch screen has to be summoned. An agent that surfaces a relevant suggestion on the home screen, in a notification, or inside an existing workflow is materially more likely to be used. The Gratitude result is a reminder that distribution, not model quality, is often the binding constraint on retention.
For teams building AI features into existing products, this argues for less time spent on conversational UI and more on identifying the two or three moments where a small, contextual surface can deliver value without requiring the user to context-switch. The home-screen widget is a useful mental model even for products that have nothing to do with mobile.
BioPharma's Quiet AI Tooling Buildout
IDBS, a Danaher subsidiary that sells cloud software to drug developers, is convening its i3 forum in Boston this month. The event itself is unremarkable industry programming, but the underlying trend is worth flagging: regulated industries are quietly building serious internal AI tooling, and the engineering practices coming out of pharma R&D are diverging from what consumer-facing AI teams are doing.
The constraints are different. Lab data has provenance requirements that make casual RAG implementations a compliance liability. Audit trails matter more than latency. Model outputs that touch GxP workflows need versioning that survives FDA inspection years later. These pressures are producing tooling patterns (deterministic prompt logging, content-addressed retrieval indexes, hard isolation between research and production agents) that the broader industry will eventually need.
Developer tools founders looking for a less crowded market than coding assistants might find more oxygen here. The buyers are sophisticated, the budgets are real, and the pain points are concrete in a way that horizontal AI products rarely are.