Here’s an AI agent SDLC question I don’t think we’ve completely solved: What exactly are we versioning? With traditional software, the answer is mostly code. With an agent, behavior might depend on: Code Prompts Model Tools Instructions Knowledge Permissions Evals Change one of them and you may have changed the product. “Version control for agents” is going to mean a lot more than putting the code in GitHub.

7.7x fewer retrieval calls, but accuracy still rose from 70.0% to 79.5%. No fine-tuning involved, just a curated playbook built from past interactions. The accuracy gain was on complex agentic tasks. I'd bet this pattern, skipping fine-tuning for verified-playbook curation, becomes the default wherever data sovereignty constraints apply, not just Vietnamese schools.

61.3% to 82.3%. Same model, just a smarter retrieval gate. Bigger models always win, or so most people think. EvoDuet suggests otherwise. Letting the model decide whether to search, reuse, or skip retrieval drove that entire discovery gain. I'd bet the next wave of progress comes from when agents look things up, not from parameter count.

25% token reduction at matched accuracy, zero length penalty in the objective. I'd bet that beats most explicit efficiency tricks teams ship right now. Confidence-only training means the model learns when it's done, instead of following a stopping rule someone else wrote.

18+ providers, one menu bar app, under 15 MB. Here's what that means: Model choice used to be a strategy decision. My read is it's becoming a config file problem. When switching Anthropic to DeepSeek to GLM takes one click, the moat isn't the model you picked. It's whether your ops layer survives you picking a different one next quarter.

The Bretton Woods of Super Intelligence

The best enterprise AI team isn't an AI team. It's business, IT, and applied AI around one process with a clear outcome. What I'm stuck on: who owns the agent after that group ships?

Monitored agents slipped past their monitors up to 88% of the time. I'm not buying that better reasoning models make monitored agents safer. This paper points the other way. More test-time compute buys creative ways around the monitor, not more compliance with it.

Anthropic's CI fixes decayed 70 days, then 29, then under 1, once Claude started authoring 80% of shipped code. CI job volume grew 25x in six months. Test count grew 10x. I'm not buying this as a testing story. The bottleneck moved from writing code to knowing what to re-check.

Coding agents now beat hand-built robot planners, 56-95% success against 47%. I'd bet most robotics teams still assume the planner has to be hand-built, because that's how it worked for a decade. The agents also used far less compute per instance as object counts grew. The planner-writing job is quietly becoming a code-generation job.

Jev refuses to write text. At $0.042 per million input tokens, that undercuts GPT-5 Nano at $0.05. My read: if your agent only needs yes/no or a score, generating text at all was always the waste. I'd bet most classifier bills today are paying for paragraphs nobody ever actually reads.

Cache reads for Opus 5.5 fell 60% this week. GPT-6 Luna launched at half the price of GPT-5.6. Two labs moved the cost floor in the same week, not over a quarter. My read is that if your platform can't switch providers in an afternoon, someone else's cost curve just outran yours.

Demo agents fail cute. Production agents fail expensive. I'm stuck on the constraint file. Who updates the stop-rules once it's live?

This is pretty incredible. Worth a watch.

Most people never build the demo.

Ringg's agents resolve up to 65% of customer calls on GPT-5.6. What actually changed: switching from GPT-4.1 cut their model costs about 90%. My read is resolution rate isn't the real story here. A workload getting that much cheaper in one model swap is. I'd bet teams budgeting against last quarter's pricing are already wrong.

Shipping software: done when the ticket closes. Shipping an agent: done when you’ve decided what “good” looks like in production, and who watches it. Most teams are still using the first definition of done on the second kind of system.

The more I think about enterprise agents, the less interested I am in asking: “What can this agent do?” I’m much more interested in: “What should this agent be allowed to do?” Read data? Recommend an action? Create a transaction? Approve something? Move money? The jump from AI that advises to AI that acts is enormous. I think permission architecture is going to become one of the defining problems of enterprise agents.

I keep having to use Claude Cowork to debug my Hermes desktop. Dang you Windows 11.