|
Hey there 👋
I checked one page first this morning: Google’s own Antigravity May preview doc. It does not bury the lede. The May agent string is deprecated and shuts down on October 5, 2026. Migrate to the September preview. That is a hard clock, not a blog tone.
My bias: the boring string in your config is the product when the vendor kills it. Fancy demos can wait. If anything in your repo still greps antigravity-preview-05-2026, today is the day it stops being theoretical.
- Ron
The Big Thing
Google kills the May Antigravity agent today
Google’s AI Studio page for antigravity-preview-05-2026 is blunt. That managed-agent id is deprecated and shuts down on October 5, 2026. The replacement is antigravity-preview-09-2026. This is not a soft sunset. Jobs, CI, and scaffolds that still pin the May string fail when the endpoint goes dark.
Antigravity here is Google’s managed agent surface: a general-purpose agent that plans, runs code, manages files, and searches inside a Google-hosted Linux sandbox. You call it through the Interactions API with an agent code string. That string is how you pin behavior. Change the string, change the loop you get.
Migration is two paths, not one slogan. If your job runs in Google’s remote sandbox and you only read final output_text or model_output, the cheap move is swap the agent id, run one representative job beside the old path, and cut over. If you parse function_call steps, drive local tools, or depend on file-edit behavior from the May surface, you have real adapter work. The September 17 preview changed local tool parameters and editing. Secondary writeups track that split against Google’s changelog. Grade them as guides. The shutdown date comes from Google’s own model page.
Why builders should care on a Monday: scaffolds and example repos still ship with hardcoded May ids. You can inherit the pin without meaning to. Run grep -r "antigravity-preview-05-2026" . across agent services, scheduled jobs, and anything that wraps the Interactions API. Treat a hit like a production incident, not a docs chore.
Ship it? ACT today. Swap the agent string on remote-output-only jobs first. Queue local-tool adapters behind a feature flag. Do not wait for a friendly error message. The page already told you the date.
Sources: Google AI Studio, Orthotropy, byteiota.
Tour de Headlines
⏳ October 9: free Gemini drops to Flash-Lite only
Update: the free-tier cut is four days out. Personal accounts with no Google AI plan keep Flash-Lite and lose Flash and Pro from October 9. AI Plus ($4.99) keeps Flash-Lite and Flash but loses Pro; those users get an email with their personal switch date. AI Pro and Ultra keep the full set. If your free stack still assumes Pro or Flash for agent work, you have a short week. Verdict: ACT if free Pro is load-bearing.
🗓️ October 14: GPT-5.5 leaves ChatGPT Work and Codex
OpenAI’s docs are still the primary. On October 14, 2026, GPT-5.5 retires from ChatGPT, ChatGPT Work, and Codex on all plans. The OpenAI API is not in that cut. If you use Codex with ChatGPT sign-in, move workspace defaults, saved settings, managed configs, custom agents, scheduled tasks, and scripts off gpt-5.5. Paid plans are pointed at GPT-6 Sol when available; Free and Go at GPT-6 Luna in the desktop app. Verdict: ACT before the 14th if anything still pins 5.5 on a ChatGPT-signed surface.
🔌 Copilot code review leaves the PR page
GitHub put Copilot code review on the REST and GraphQL APIs (Oct 2, GA for Pro through Enterprise). You can request a review from scripts, workflows, and internal tools, and set effort per request. Balanced is now the default effort level (effective Sep 28 for Default; explicit Lite stays). If CCR only lived in the pull request UI for your team, the API is the path into CI. Verdict: DEPLOY if you already trust CCR and want it off the browser.
|
Sponsor
Your calls leave a signal. Read it.
RapportScore measures how your team communicates on real calls and coaches them on it. It scores what happened on the call, using human signals plus AI, so your people get better at the conversation itself, meeting after meeting.
See your team’s score →
|
Tool of the Day
🧬 Google RRSI
Regularize the harness loop so it stops memorizing the test.
Google Cloud AI Research open-sourced RRSI: Regularized Recursive Self-Improvement of agent harnesses. The model stays frozen. The harness (prompts, control flow, tools, memory, sub-agents) can change. Unregularized evolution often posts big gains on the evolve set and then flatlines or loses on unseen tasks. RRSI caps edit budgets, shrinks them over rounds, and uses a critic that rejects benchmark-specific tricks. Reported results: up to 14.1 points on evolve tasks, up to 4.7 on unseen benchmarks, and about 30% fewer runtime tokens than unregularized evolution. Code is Apache 2.0. Treat vendor numbers as vendor numbers. The useful idea is simple: if you let an agent rewrite its own harness against a fixed eval, add regularization or you bought a memorizer.
Open the RRSI repo →
Worth a Click
Three clocks on one desk. Google kills a managed-agent id today. Free Gemini loses Flash and Pro on the 9th. GPT-5.5 leaves ChatGPT Work and Codex on the 14th. None of that is a keynote. All of it is config.
My bet: the teams that stay calm this month are the ones who treat agent strings and model pins like dependency versions, with greps in CI and a named owner. Everyone else will discover the pin when the job turns red. The demo can wait. The string cannot.
- Ron
|
