Topic
Models and plans aimed at agentic software engineering, shell/tool use and end-to-end issue fixing.
Block open-sourced the desktop app its own teams use to drive agents. The design claim underneath it is that active context should be inspectable, not archaeology in a prompt log.
Reality
Salesforce says builds, tests and reviews validate the diff, not the requirement. Its answer was a specification gate upstream of the code, plus a rule about which questions agents may answer.
Reality
OpenAI is rolling browser and app control out to customers while Anthropic pushes Claude Cowork. The open question is not capability but which credentials the software gets, and who reviews the log.
Reality
The maintainer removed holistic "looks good" verdicts, moved planning to a separate cheap model, and left the gate closed when verification cannot run. Breaking for embedders.
Reality
A denial of a Bloomberg report on SpaceX buying Cognition leaves the compute relationship unaddressed, and leaves buyers choosing coding tools from a shrinking independent field.
Reality
Bloomberg says SpaceX approached Cognition. Its CEO says no talks happened. Either way, the pattern is compute traded for developer data, and that reprices any single-vendor coding bet.
Reality
OpenAI merged an async developer-message tool into the public Codex repository, so the agent no longer blocks on your answer. Nothing in the change stops it from coding past your decision.
Reality
JetBrains traced a model through fifteen C# refactoring tasks and found it simulating structure with sed, git and the compiler. Wiring in Rider's real engine cut median time by 83%.
Reality
ProdCodeBench builds tasks from real assistant sessions in an industrial monorepo. Its authors report that models using validation tools more heavily solve more, which argues for scoring tool discipline.
Reality
A Paris start-up says its AI agent turns plain English into running quantum simulations. Three test cases, one of them a deliberate failure, is the whole evidence base so far.
Reality
A dev.to post argues coding agents fail because abandoned decisions and standing conventions come back with identical confidence. The proposed fix is a schema with explicit precedence.
Reality
Origin keeps GitHub as the source of truth. That is both the reason engineering orgs can try it this quarter and the reason it is not yet an exit.
Reality
The seed round funds agents that map and document mainframe code before rewriting it. Comprehension is the sellable half; the two-month timeline is still the company's own number.
Reality
The distro's manual says agents are first-class citizens but it will not pick a favourite. Lazy-loaded stubs make that neutrality cheap, until a crash needs a default agent.
Reality
RuntimeWire says it drove an agent thread in Cursor Desktop from a terminal using a gated feature that appears in neither the CLI guide nor the changelog.
Reality
JetBrains reports 90% of professional developers use coding agents weekly. The number that matters for tooling bets is not adoption but conversion, where Claude Code is far ahead.
Reality
A dev.to writeup argues long sessions degrade structurally, not linguistically, and proposes research/plan/implement phases with a hard context clear between each.
Reality
A dev.to tip argues agent pipelines need an adversarial verifier from a different model family plus randomized human audits, because same-model review exploits a documented self-preference bias.
Reality
OpenAI's Codex update adds computer use on Windows and remote control, per a dev.to write-up. That changes what an engineering team hands off, and how much a host machine can be trusted.
Reality
Investors at UBS's Venture Capital Summit expect surviving SaaS firms to drop seat-based economics. Product leaders now have to defend renewals on outcomes they can actually measure.
Reality
Mads Thines has shipped an open-source memory layer for coding agents that starts on local disk. The design bet is that a record of past mistakes should be inspectable and portable.
Reality
A thirty-run experiment finds agents followed long, buried rules anyway. The two failures came where the file contradicted the repository, not where it was long or buried.
Reality
Anthropic's own study found users approved 97% of prompts and caught 13.6% of harmful actions. From August 14 the click stops being the safeguard, and deny rules become the job.
Reality
Stephen Cresswell rebuilt a 2012 test library in about a day with Claude Code. The transferable part is the commit discipline: a phased issue, formatting kept apart from behavior, tests off limits.
Reality
A developer benchmarked his local agent runner before tuning it and found his instinct was wrong twice over: the event loop was idle, the hardware was half used, and the real work was serialized by a rule.
Reality
Platform teams planned capacity by headcount. If tenancy demand now scales with changes in flight, a namespace per developer is oversubscribed on arrival.
Reality
A dev.to comparison of OpenSpec and GitHub Spec Kit argues chat degradation is structural. The interesting part is not the diagnosis but how differently the two tools file the cure.
Reality
A dev.to write-up splits agent work into prompts, context and harness. The interesting part is the harness: tool execution, permissions, validation and recovery, all of it code you own.
Reality
Sol-Luna's supervisor could have split independent modules across parallel workers. Given a free choice across six benchmark runs, it kept the work for itself every time.
Reality
A runtimewire columnist argues Anthropic won the industry's centre of gravity through Claude Code and then spent the credit down. The growth half of that case has numbers. The decline half does not.
Reality
One developer's answer to unverifiable agent self-reports: rules compiled into code, a verdict that is a pure function, and no model credential anywhere in the verdict path.
Reality
Z.ai says its new model tops CyberGym and leads open-source models on Terminal Bench 3.0. The weights go to Hugging Face within two weeks, which is the part security teams should read twice.
Reality
The AICPA's change-management criterion asks whether changes are authorized, tested and approved, not whether a second engineer clicked approve. Amp says its auditors agreed.
Reality
Z.ai says its 743B-parameter GLM-5.3 hits 34.5% on its own code bench using 22% fewer output tokens than GLM-5.2. The weights are still two weeks out.
Reality