Notes from inside the work.

Post-GPU is usually three different ideas sharing one headline, and all three are really about moving data. Where is your team's infrastructure planning on that curve? #EricFieldNotes

A company evaluating local AI should compare delivered work, not model ownership in isolation. A smaller local model, a hosted frontier API and a dedicated cluster have different privacy, latency, operating and quality profiles. Define the …

It is easy to buy more agent seats. It is harder to create the owner time needed to decide what those agents should build and whether their work is acceptable. Picture ten parallel patches blocked on one unresolved tenant policy. A model c…

If ten agents wait on one tenant-policy answer, the next license is not the highest-value purchase. Give an accountable owner a short decision queue, evidence packets and a response window. Then add lanes only as accepted work grows. #EricF…

I overwrote a running script, and bash ran a command nobody wrote. Never edit a running script in place. #EricFieldNotes

A rule in a markdown file is useful guidance. It is not the same control as a pre-tool gate, a post-tool readback or a capability limit outside the worker. This film separates those layers. Agent instructions can explain intent. Hooks can …

An agent rollout can make the activity dashboard look fantastic: more runs, more diffs, more test summaries. The business may still wait longer for a correct release. Suppose three agents implement three versions of the same tenant excepti…

A team doubles agent runs and PRs; release lead time rises because policy decisions wait on one owner. Count accepted changes, review age, integration rework and reversals by work class. More activity is a hypothesis about productivity, not…

A model's confidence is the model's opinion of itself. It shouldn't be what moves money. In the systems I build, the model proposes and something else decides. #EricFieldNotes

My AI agent coordination project ended BLOCKED, and I am publishing the scorecard anyway. The question was how coding agents on Claude Code, Codex, Grok Build and Muse can message each other, wake each other, follow instructions and be hel…

Hook support is not one universal property of an agent product. The meaningful question is whether your callback runs inside the exact loop that can cause the effect. A team registers a pre-tool rule for local shell commands and assumes th…

A pre-tool hook may see local shell calls while a hosted worker owns another loop. Inventory routes, inject a harmless forbidden action, capture the hook receipt and read the target. Deny at the credential boundary when the callback is out …

My main agent sessions re-read about half a million tokens on every turn. I measured it across 35 days and 5,835 transcripts. Every request re-reads the whole conversation, and in my long-running sessions the median turn carried roughly 45…

“It loads” is not “it serves.” Offload, cold start and queueing can dominate. Keep the placement that meets the whole contract. #EricFieldNotes

A router may choose a valid tool bundle and still send the worker into a different permission set. The gap between selection and dispatch is a normal time-of-check/time-of-use boundary. In a disposable fixture, give the router a read-only …

A router picks a read-only tool bundle. Before dispatch, the same name gains a write route. Fingerprint schema and scope with the choice; change one harmless field in a test. Dispatch must reject stale authority and log any fallback. #EricF…

An agent can implement a change and write its own tests in one run. That is useful speed. It is also one interpretation producing both the work and the evidence. Let the worker write useful tests. Keep the final verdict independent. #EricF…

Kimi K3's model card lists two point eight trillion total parameters, one hundred four billion active per token and native four-bit weights. Sparse activation changes compute. The full expert set, memory capacity, movement and concurrent se…