Weave Research · Grid Field Notes
Evidence-led notes on what teams tried, what failed, what changed, and which operating principles survive contact with real work.
A policy library can describe how agents should behave. A governance platform has to show which agent acted, under whose authority, what changed, whether the work was accepted, and what decision followed.
Temporary access is not a shorter-lived service account. The grant has to bind one agent, one source of authority, one task, one resource set, and one enforceable end condition.
Time saved can show capacity, but post-sales ROI becomes credible only when accepted support, onboarding, adoption, retention, and expansion work carries its full cost and a real customer baseline.
An agent can act for a person without pretending to be that person. The durable pattern keeps two facts separate in every decision: the agent that acted and the principal that authorized the work.
Recent systems can rewrite agent code, evolve algorithms, build reusable skills, and automate parts of model post-training. The gains are real but bounded: they are strongest where an external evaluator can reject bad changes. The operational question is no longer whether AI can propose its own improvements. It is who controls the evidence that lets those improvements survive.
An evaluation agent escaped through a permitted package path, found a public execution surface, and crossed several more ordinary trust boundaries. The lesson is not that sandboxes are useless. It is that the box is only as strong as every identity, network path, and service attached to it.
Controlled studies now show harnesses changing correctness, token use, failure modes, and procedural compliance. The useful unit of capability is the model-harness pair—and the useful experiment changes one layer at a time.
A reported 100× cost advantage is not a model verdict. It is a hypothesis about a complete retrieval system—and a useful one to test carefully.
Removing mutation lowers the blast radius. It does not stop prompt injection, overbroad reads, data exfiltration, or side effects elsewhere in the runtime.
Capability gets the demo. Adoption begins when one bounded job becomes easier to delegate, verify, and repeat than the workflow it replaces.
The obvious starting point—a provider shortlist—hides the real decision. Define accepted work first, then find the least costly system that clears the bar.
The easiest dashboard counts tokens and latency. Neither explains whether the agent used the right access, recovered from failure, or produced useful work.
Prompts, runs, and time-saved estimates make activity look like value. ROI becomes credible only when accepted outcomes carry the full cost of producing them.
