The Agent Gets a Job Description: DeepSeek's Always-On Coder, a Sunset Date for Jules and Big Tech's $3 Trillion Tab

Sunday, 1 November 2026 · Global AI News Daily Ledger

Coverage window Since 31 October, 18:47 IST | Compiled 1 November, 18:44 IST

The week the agent got a job description. DeepSeek open-sourced an always-on GitHub agent built to take as little human attention as possible; Google announced the sunset of Jules, which goes offline on 31 December, to concentrate its coding agents in Antigravity; and the subscription war moved to quotas - Anthropic has dropped the five-hour clock for verified Ultra accounts, while OpenAI said ChatGPT and Codex will merge into one metered pool from December. Security kept the drumbeat: Chainalysis named 'offensive agentic techniques' as its intruder, and researchers showed a shared ChatGPT link could take over an account. The money reached $3 trillion in future commitments, and the referees moved in: xAI must face the NAACP's Clean Air Act suit, the UK opened a probe into Meta's terms, and Canada joined EuroHPC days after buying into Cohere.

1. DeepSeek open-sources dsh-superagent, an always-on router-worker coding agent

Reporting. DeepSeek's Harness team released dsh-superagent on 1 November: an MIT-licensed research prototype of a proactive coding agent for a GitHub repository, shipped with a design post ('Exploring Super-agent Design'). A long-running Router session listens to GitHub (and optionally Feishu), decides what counts as authorized work, and dispatches each task as a Ticket to an isolated Worker session that returns results - often as a pull request - to the original conversation. The stated design rule: take as little human attention as possible; the agent confirms once, stays quiet while working, and speaks again only with a result, a question or a failure. Proactive work is gated behind accept modes (off, observe by default, auto). DeepSeek's own simulation reports p95 task completion of 15.7 minutes versus 29.5 for a single-session baseline - self-reported, on synthetic message streams, with no real code-quality measurement. It runs locally as a Harness profile with full OS-user access, so isolation is on the operator.

Analysis. The third major open agent artifact from DeepSeek this month lands squarely on Claude Code's turf, and the architectural claim is the interesting part: human attention, not model capability, is the bottleneck, so the unit of design shifts from the session to the dispatch layer. The attention economics are right; the benchmarks are marketing until measured on real work.

DeepSeek · SCMP / MarkTechPost / The Next Web · 1 Nov 2026 · GitHub repo · MarkTechPost report · SCMP report · The Next Web report

GitHub repo · MarkTechPost report · SCMP report · The Next Web report

2. Google sunsets Jules and consolidates its coding agents into Antigravity

Reporting. Google announced on 29 October that Jules, its asynchronous coding agent launched in December 2024, winds down: sessions stop accepting new work on 11 November, the product turns read-only, and the service goes offline on 31 December. An export tool covers account settings, environment variables, scheduled tasks, connected repositories, saved memories, setup scripts and full session histories, but there is no automatic migration into Antigravity, the agent-first platform Google introduced with Gemini 3 and upgraded at I/O. The same day, Antigravity launched a Plugin Marketplace for app and CLI - one-click bundles of skills, MCP servers, rules, hooks and custom agents, with launch partners including Atlassian, Figma, Stripe, Datadog, Databricks, Canva, CrowdStrike, Dropbox and Unity. Google has not said when Jules-style background tasks arrive in Antigravity. Missed by the 29-30 October editions; dates noted.

Analysis. Consolidation, not retreat: Google is betting the agent platform beat the standalone background agent. Teams that built routines on Jules have weeks to rebuild them - the cost of riding an experiment.

Google · The Verge · 29 Oct 2026 · Jules changelog · The Verge report

Jules changelog · The Verge report

3. The subscription war moves to quotas: Claude drops the 5-hour clock, OpenAI merges ChatGPT and Codex limits

Reporting. Anthropic renamed its consumer tiers on 31 October - Max 5x becomes Max, Max 20x becomes Ultra, prices unchanged - and removed the 5-hour rolling session limit for verified Ultra accounts, so long coding sessions can run until the weekly allowance (still 10x Pro) runs out; most existing subscribers already qualify. The same week OpenAI doubled Pro-tier Codex and ChatGPT Work allowances, then announced that ChatGPT and Codex will share one unified, token-metered quota pool from December, with credits visible and purchasable - and began capping the number of paid ChatGPT accounts per device and payment card to curb sharing. Anthropic subscription lead Michael Overell apologized for the confusing 20x naming; the moves land weeks before Anthropic's expected IPO, with bond investors reportedly told to expect a BB rating.

Analysis. Quota generosity is now a competitive front: Anthropic sells uninterrupted time, OpenAI sells a single metered pool. Both are really about agents - long autonomous runs are what break 5-hour windows and separate app quotas.

Anthropic · OpenAI Help Center / Windows Report / 36Kr · 30 Oct-1 Nov 2026 · Claude Ultra plan (help center) · Windows Report · OpenAI usage limits (help center) · OpenAI card and device limits

Claude Ultra plan (help center) · Windows Report · OpenAI usage limits (help center) · OpenAI card and device limits

4. Chainalysis names 'offensive agentic techniques' as its intruder

Reporting. Chainalysis disclosed on 30 October that an external actor used offensive agentic techniques to gain unauthorized access to its systems and internal tools - among the first times a breached vendor has put agentic AI on the intrusion-method line of an official disclosure. The company says contact-level data (user IDs, names, email addresses, organization role labels) was accessed; it found no evidence that investigation content, transaction or query results, cleartext passwords, MFA secrets or session tokens were touched, and states the incident is unrelated to the recent Binance theft it helped attribute to North Korea. Chainalysis says it is itself using large language models for a broader security review - AI on both sides of one incident.

Analysis. Coming the same week as CrowdStrike's VAPOR PANDA report and the Cursor extension backdoor, this is the normalization of the agentic breach disclosure: 'agentic techniques' is becoming a standard line in the incident template, from the firm that investigates crypto crime for a living.

Chainalysis · BleepingComputer / The Block · 30-31 Oct 2026 · Chainalysis disclosure · BleepingComputer report · The Block report

Chainalysis disclosure · BleepingComputer report · The Block report

5. A shared ChatGPT link could hijack an account: Parameter Labs' sandbox escape chain

Reporting. Parameter Labs disclosed on 28 October a complex attack chain against ChatGPT's interactive HTML previews (the artifacts pattern): the sandboxed iframe could drive the MathLive math library's virtual keyboard in the parent page via postMessage, and message-handling code using innerHTML turned that into a sandbox escape and account takeover, triggerable through a shared chat link. OpenAI is working to strengthen the sandboxing. Simon Willison amplified the writeup on 31 October, stressing that HTML in these previews must stay tightly sandboxed because untrusted code can otherwise reach OpenAI APIs and private data.

Analysis. Agent-era UI keeps recreating the web's old mistakes: a sandboxed preview that can message its parent is an XSS problem wearing new clothes. The shared-link vector makes it social, not just technical.

Parameter Labs · Simon Willison · 28-31 Oct 2026 · Parameter Labs writeup · Simon Willison's analysis

Parameter Labs writeup · Simon Willison's analysis

6. Bloomberg: Big Tech now carries over $3 trillion in future AI spending commitments

Reporting. Bloomberg calculated on 30 October that disclosed commitments for leases, construction, energy and equipment tied to the AI buildout now exceed $3 trillion across Alphabet, Meta, Microsoft, Amazon and Oracle - separate from capex already spent. Alphabet alone accounts for more than $1 trillion in contractual commitments, nearly eight times a year ago; Meta tops $800 billion, almost half in data-center leases that have not yet started, on schedules up to 30 years. About $1.3 trillion is future lease payments and $1.9 trillion other purchase commitments. Caveats: disclosures are not standardized, some Meta commitments cover Reality Labs hardware, and Alphabet and Amazon include content licensing. The tally landed as the Nasdaq fell nearly 3% for a third straight weekly loss and Intel posted its worst single-day decline in over four decades.

Analysis. This is the liability side of the AI trade finally getting a number - commitments that survive a sentiment turn. Combined with OpenAI trimming its round (31 October edition), the question has shifted from 'who can raise' to 'who must pay'.

Bloomberg · Yahoo Finance / The Next Web · 30 Oct 2026 · Bloomberg report · Yahoo Finance (syndication) · The Next Web report

Bloomberg report · Yahoo Finance (syndication) · The Next Web report

7. xAI must face the NAACP's Clean Air Act suit over its Memphis data center

Reporting. A federal judge denied xAI's motion to dismiss the NAACP's Clean Air Act lawsuit on 30 October, rejecting the claim that the roughly 35 combustion turbines powering the Colossus data center in Memphis are exempt 'mobile sources'. The suit alleges xAI never obtained preconstruction or operating air permits for the turbines. The ruling lets the case proceed to discovery; Earthjustice, which represents the plaintiffs, called it a denial of xAI's attempt to escape the permit regime. It is the second procedural loss for the company in the case in two weeks, after a judge threw out a DOJ attempt to block the community suit on 19 October.

Analysis. The legal frame for data-center power is being set in Memphis: if turbines that move are still sources that need permits, every fast-tracked gas-powered AI campus inherits the problem.

TechCrunch · CNBC / Earthjustice · 30-31 Oct 2026 · TechCrunch report · CNBC report · Earthjustice statement

TechCrunch report · CNBC report · Earthjustice statement

8. UK competition authority opens investigation into Meta's commercial terms

Reporting. The UK's Competition and Markets Authority opened an investigation on 30 October into Meta's proposed commercial terms for businesses and publishers using its platforms, examining whether the terms could impose unfair conditions - including around content and data use for AI. The publishers' trade body PPA welcomed the probe; Euronews frames it as part of a global pushback against the terms. The investigation is at an early stage and no finding has been made.

Analysis. AI training value is now a competition-law question: regulators are treating what platforms demand in their small print as market power, not contract detail.

GOV.UK · Euronews / PPA · 30 Oct 2026 · CMA announcement · Euronews report · PPA statement

CMA announcement · Euronews report · PPA statement

9. Canada joins EuroHPC as its 39th participating state

Reporting. Canada became the 39th participating state of the EuroHPC Joint Undertaking on 30 October, announced at the EU-Canada summit in Montreal - the first non-European country at full participating-state level, giving Canadian researchers and companies access to EU supercomputing and quantum resources, including a fast lane to the JUPITER exascale system. The move follows Canada's up-to-C$1.6 billion equity investment in Cohere earlier the same week (30 October edition).

Analysis. Sovereign AI week for Canada ends with compute access: equity at home, federation abroad. Middle powers are buying their way into every layer of the AI stack rather than choosing one.

EuroHPC JU · European Commission / POLITICO · 29-31 Oct 2026 · EuroHPC announcement · European Commission · POLITICO report

EuroHPC announcement · European Commission · POLITICO report

10. Anthropic starts enforcing its no-training terms, and researchers say legitimate work is being blocked

Reporting. Users reported on 31 October-1 November that Anthropic is blocking accounts it suspects of using Claude outputs to train or distill rival models, enforcing terms updated on 8 October that ban training on outputs. Developers and researchers complain the enforcement also catches legitimate troubleshooting and non-malicious research; FourWeekMBA and Hacker News threads document affected users. Anthropic has not published enforcement criteria or an appeals process. Sourcing note: this story rests on user reports and secondary coverage, not an Anthropic announcement.

Analysis. Every lab is now policing the output-to-weights pipeline, but enforcement without visible criteria turns a distillation ban into a research-chilling guess. Watch for a formal appeals process - or a competitor marketing one.

FourWeekMBA · Hacker News · 31 Oct-1 Nov 2026 (user reports, unverified by Anthropic) · FourWeekMBA report · Anthropic usage policy update (8 Oct)

FourWeekMBA report · Anthropic usage policy update (8 Oct)

Sourcing notes

Stories 2, 5 and 6 were published 28-30 October and missed by that week's editions; their original dates are noted in line.
dsh-superagent's performance figures are DeepSeek's own, measured on synthetic message streams; no independent code-quality benchmark exists yet.
Story 10 rests on user reports and secondary coverage; Anthropic has not announced enforcement criteria, and the underlying terms date to 8 October.

Compiled 1 November, 18:44 IST. Exact source links above.

TechSambad