The clearest change this month is what people chose to publish: not "look what the model did", but what it cost, how often it failed, and who checked it.
Simon Willison shipped a release of a library he has maintained for years and wrote down the bill: $149.25 of API spend, 37 prompts, 34 commits, most of the code written by Claude Fable. The detail worth stealing isn't the total. It's the breakdown. The main session was $141.02. The five review agents he ran alongside it cost between $0.32 and $2.40 each, and one of them found a connection-poisoning bug in delete_where() that had survived his own testing. He then had GPT-5.5 review Fable's work, which surfaced further transaction edge cases. His own conclusion was that he should have pushed more of the work onto cheap subagents than he did.
Cross-model review costs almost nothing and catches things the author doesn't. He calls the practice "weirdly superstitious" and says it works anyway.
Where the bottleneck moved
Verification, and specifically human attention as a scarce resource.
Addy Osmani made the sharpest version of the argument in Own the Outer Loop on 15 July. The inner loop (investigate, implement, verify) goes to the agent. The outer loop (quality signals, the ship-or-block verdict, and answerability for that verdict) stays with a human. He puts three numbers behind it: 42% of committed code now AI-generated or AI-assisted per Sonar's 2026 data, 73% of people in a Wharton study accepting incorrect AI output and feeling more confident afterwards, and an Anthropic trial where engineers using AI comprehended the resulting system 17 points worse, 50% against 67%.
Those numbers all point the same way. Throughput isn't the constraint. Knowing whether the output is right, and being able to say why you shipped it, is.
Hamel Husain tested the obvious fix and found it partial. In Do Automated Evals Work? on 11 July he ran automated trace-analysis tools against real production data. They do spot issues humans miss, and they slot cleanly into existing trace-review workflows. They also miss anything that needs domain expertise or taste, and they have no good mechanism for learning from human corrections. His recommendation is iterative use with a human in the loop, which is the same shape as Osmani's outer loop arriving from a different direction.
What is contested
Whether coding is finished. Boris Cherny, who created Claude Code, said on Lenny Rachitsky's podcast that coding is "largely solved" for the kind of work he does, and that he has not hand-edited a line since November 2025 while shipping 10 to 30 PRs a day. Osmani takes that on directly in Earning taste and judgment (14 July): "the last mile, edge cases, architecture, taste, is the whole game." He backs it with labour data rather than vibes: recent-graduate unemployment at 5.6% in March 2026, underemployment at 41.5%, junior tech titles down 34% since early 2020 against 19% for senior roles. These two aren't in dialogue, and Cherny's remark is months older, so treat it as two positions rather than an exchange. The gap between them is really about which part of the job you're counting.
Whether better models are better tools. Armin Ronacher's Better Models: Worse Tools (4 July) is the most useful counterintuitive finding of the month. Newer Claude models, on tool schemas that do not match the shapes they were trained against, invent extra fields in the call. The payload is right, the keys around it are hallucinated: requireUnique, matchCase, oldText2. He saw roughly a 20% failure rate in one session with Opus 4.8, halved it by stripping thinking blocks from history, and eliminated it with strict tool invocation mode. Older Anthropic models didn't do it. The Codex models he tested didn't either.
The generalisable claim is the one to keep: tool schemas are not neutral. If your tool shapes are unusual, the model's training priors are a force acting against you, and that force gets stronger as models get better at everything else.
Worth trying
- Run a review agent on a cheap model against your own diff, before a human sees it. Willison's cost data says the marginal price of this is a rounding error against the main session.
- Review across models, not just across people. One model writes, a different vendor's model reviews. It found real transaction bugs in a library maintained by an expert.
- Turn on strict tool invocation if your provider has it, and add tolerant key filtering or parameter aliases if it doesn't. Cheaper than arguing with a prior.
- Write down who owns the verdict for each class of change, and what evidence they need to see. That's Osmani's accountability contract, and it's a document, not a tool.
What they adopted themselves
MCP got interesting again because it got simpler. The 28 July spec revision drops the stateful bidirectional session for a request-and-response core and makes servers OAuth 2.1 resource servers. Willison, who had drifted away from the protocol, built three implementations in a week: a CLI for probing servers, a plugin exposing SQL over MCP, and a client for his own LLM tool. His stated reason for coming back is the security story, not the ergonomics: he now plans to use MCP specifically for sensitive applications, where a stateless, properly-scoped server is easier to reason about than handing an agent a shell.
The other adoption pattern is packaging. The recurring line out of the AI Engineer World's Fair, in Latent Space's write-up on 14 July, came from Google DeepMind's Philipp Schmid: "Agents are just files. We write markdown files to extend capabilities." Every major agent platform now has a skills mechanism, and the same event pushed loop engineering, the outer-loop supervision layer, as the second of its five trends.
Watching
Whether the constraint layer turns out to be deterministic or just bigger models.
The revival of ontologies is the interesting version of this question. Latent Space's 30 July piece documents teams reaching back to semantic-web machinery to bound what a probabilistic agent is allowed to conclude, with Neo4j building ontology-backed semantic layers under "thin agents". The counter-argument is in the same piece and is the one that killed the semantic web the first time: somebody has to maintain the ontology, and Emre Sokullu's view is that capable enough models eventually stop needing it. Nobody has evidence either way yet. Whichever side wins decides whether the outer loop is a schema or a second model.