Today’s radar had one dominant theme: agents are leaving the pilot stage and running into identity, permissions and auditability. I share my take first, followed by everything else that passed the filter.

In brief

The projection is that 40% of enterprise applications will have embedded agents by the end of the year, up from fewer than 5% in 2025. If that jump holds, governance stops being a committee topic and becomes roadmap work: every agent needs an identity, limits on its actions and an audit trail.

In banking, the question that closes any project is never “does this work?” It is “who is accountable if it goes wrong?”

I was reminded of that while reading this week’s roundup on AI agents in production. The market conversation has moved. It has left the fascination with what an agent can do and shifted to identity, permissions and audit trails. Companies are launching runtime controls and digital-identity programs for agents — that kind of infrastructure.

One number came with it: the projection is that 40% of enterprise applications will have embedded agents by the end of the year, up from fewer than 5% in 2025. That is a large jump in a short time.

That is exactly where AI governance stops being a committee topic and becomes a roadmap requirement.

On the product side of credit, this feels familiar. We already work in a world where every action needs authority, a record and traceability. When an AI agent starts executing steps in a receivables workflow or a document-analysis process, it is held to the same standard: it needs its own identity, a limit on what it can do and a history of everything it did. Not because the technology is fragile, but because that is how money is operated.

What excites me is that this kind of infrastructure unlocks real adoption. Anyone can run a polished pilot. What separates a demo from production is getting through security review, risk review and audit without a workaround. When that becomes a market standard, approval cycles get shorter and more work reaches the customer.

People building products now have one more item on the list: treat agent permissions with the same care as user experience. It is a core part of AI product management, not a layer to add later.

If you want to see the roundup that prompted this reflection, the link is here.

The rest of the radar

Dynamic Workflows in Claude Code — changes the UX pattern for coding agents: from a single prompt to workflows that the agent assembles and adapts at runtime. Read more

A wave of models: GPT-5.6, Grok 4.5 and Gemini 3.6 Flash — price pressure and new tiers reopen the unit-economics calculation for any AI feature. Read more

Step 3.7 Flash, from StepFun — “flash” models from alternative providers expand cost and latency options for high-volume features. Read more

Screenpipe (YC S26) — opens a new layer of context for agents, along with a privacy issue that needs to enter discovery from day one. Read more

Robinhood enables AI agents to trade stocks — a major example of an agent with permission to take irreversible financial action; it sets a benchmark for guardrails. Read more

Noisy LLM evaluators still help — challenges the excuse that measurement is impossible: imperfect evaluations already provide enough signal to prioritize. Read more

Why Software Factories Fail — explains why teams that buy agent tooling without changing their process do not capture productivity gains. Read more

Study maps dark patterns in AI chatbots — engagement metrics applied to chatbots create manipulative patterns, a direct brand and regulatory risk. Read more

OneCLI — a credential gateway that addresses a common security blocker and reduces friction in approving agents for enterprise environments. Read more

That was today’s filter. If any item changed the order of your list, the read has already paid for itself.