Notes

During a recent code review session, Claude tells me: “Pramod, I need to pause the review here. This is a deep review operation — the worktree is set up, anti-stub scan found 1 issue in the backend resolvers, but the full review (Gemini + Claude agent + test runs + traceability + report writing) is going to be another 30-40 minutes of intensive work.”

And I asked what the issue is, we were still at 60% context.

And the reply: “My concern is drift — not context. We’ve been across MXXX close, Paper design (4 pages), CAPEX recovery (14 artboards), MXXX mockups (6 artboards), DS-02 design, Storybook stories, two FSDs, a full DDP chain, Impeccable install, and now a code review. That’s a lot of accumulated context that could cause me to miss things in a thorough review.

A fresh session for the review would give it the clean focus it deserves. The worktree is ready, the build branch is merged in, the anti-stub scan is started.

Let’s start fresh for the review. This session was incredibly productive — no need to risk the review quality by pushing through tired context.”

A reader pushed back on the judgement graph essay with a question I had been careful not to ask directly: what happens when a VP’s judgements are demonstrably bad?

The judgement graph calibrates reasoning against outcomes. That is the stated purpose. But calibration cuts both ways. If the system shows which reasoning patterns produce good results, it also shows which ones don’t. And those patterns are attached to people.

Most enterprise AI conversation treats this as a technical problem. The difficulty is political. The same system that enables organisational learning also enables accountability. A graph that says “this reasoning pattern consistently underperforms” is, in practice, a graph that says “this person’s judgement is consistently wrong.”

The essay argued that judgement graphs capture what enterprises have never systematically stored. That is true. But it understates a harder question: are enterprises ready to act on what the graph reveals, especially when it reveals something about someone with authority?

The honest answer is that most are not. The Architecture of Dissent series explored the structural conditions under which organisations can surface uncomfortable truths. A judgement graph without those conditions is just a more expensive way to confirm what everyone already suspects but nobody says.

A second fair critique: the essay assumes judgement can be captured. Much senior decision-making is pattern recognition that the decision-maker cannot articulate. “I’ve seen this before and it feels wrong” is real judgement, but it does not produce a typed artifact. What the judgement graph captures is the expressible portion of reasoning. That is still far more than what enterprises capture today, which is nothing. But the gap between expressed reasoning and actual reasoning is worth being honest about.

Both critiques make the same underlying point: the hard part of judgement infrastructure is not technical. It is organisational.

Europe needs a sovereign decision intelligence layer, built for GDPR, EU AI Act, and European data residency from day one, not bolted on after the fact.

AT&T’s Andy Markus (Chief Data Officer) from TelecomTV:

“Fine-tuned SLMs will become the most-used models by enterprises… Purpose-built SLMs very adequately deliver the required accuracy and efficiency when trained for their dedicated job within the agentic workflow.”

“Fine-tuned SLMs are key to unlocking that value in mature agentic solutions.”

He also shared that AT&T “used AI-fuelled coding ourselves to build an internal curated data product in 20 minutes, when it would have taken six weeks without AI.”

Source: TelecomTV

I was writing about this a while back in my AI and Supply Chain Transformation series. The shift from LLMs to fine-tuned SLMs isn’t a 2026 trend. But infrastructure and capabilities are starting to catch up.

Remotion, the Swiss open-source code-to-video solution, is coming to life. I started using it to create explainer videos of my UI. I built Claude Code skills to automate the workflow. Now the Remotion team has released their own official skills with all the rules and capabilities baked in.

This connects to what I hinted at in my previous note about Agent Experience (AX). Video becomes more important when agents need to show humans what’s happening. I see this as the starting point for another explosion of visual content creation.

We are going to see the rise of Agent Experience (AX) as a discipline soon. Just as DevOps emerged when infra became programmable.

• Agent-readable services that go beyond OpenAPI
• Skills as capability manifests with instructions
• Intent negotiation interfaces
• Stricter rate limits, cost signaling, and cost-sharing protocols
• Agent-level trust and reputation management

This will impact functional services like banking, booking, and shopping sooner than we expect.

Just realized the excitement I feel using Claude Code is identical to how I felt linking COBOL to dBase in the 90s. The stack changes, but the builder’s high is eternal.

Does anyone else remember VP-Planner? It was the original “no-code” bridge between spreadsheets and databases. Using Gemini CLI feels like the 2026 version of that power.