Morning digest: DeepSeek agent, Codex, and chatbots

Morning digest for Sunday, Nov 1. DeepSeek released the code for "Super-agent," an always-on coding agent that watches a GitHub project and picks up work when a teammate asks, as Chinese labs push into the market for tools like Claude Code. Also: OpenAI is giving some Codex users free "banked" resets of their usage limits, and here's how to use one; DeepSeek's web search and page-fetch tools started charging per call; The Washington Post found chatbots leaning on a small AI-built voter guide two days before the U.S. election; Google made its Computer Use tool for Gemini generally available; a judge let the NAACP's air pollution lawsuit against xAI's data center move forward; and AI agents are now running the Allen Telescope Array in the search for alien signals.

1. DeepSeek open-sourced "Super-agent," an always-on coding agent that waits in your team's GitHub for requests. DeepSeek's Harness team, which builds the software scaffolding around its models, published dsh-superagent-prototype on GitHub on Saturday under the MIT license, along with a long design article (English version on X). The agent is tied to one GitHub repository and, optionally, Feishu, a workplace chat app popular in China. When someone who's allowed asks it for work, it opens a work item, does the job in its own separate session and folder, and answers in the original conversation, for example with a pull request (a proposed code change). A web panel shows what needs a person. The README's guiding rule is to "take as little human attention as possible, and give as much value as possible," so it acts only when asked or when a rule the team wrote allows it. The idea behind it, in DeepSeek's words: "a person can usually follow only 3 to 5 active sessions at once," so a single task is limited by the model, "but multi-tasking is limited by human attention." Its design splits the work in two: a lightweight "Router" that talks to people and hands out jobs, and several "Workers" that do the jobs in parallel. In DeepSeek's own simulated test (about 300 chat and CI messages, about 20 of them real work requests, with the coding itself faked), Super-agent finished tasks in a median of 12.4 minutes, versus 20.4 minutes for a single agent session that was allowed to hand off work (13.8 when its prompt pushed it to delegate). It also used roughly four times as many cached tokens (23.2 million versus 6.1 million per run), and the team calls the results "indicative only." Bloomberg reports that DeepSeek is aiming to challenge Anthropic's Claude Code. DeepSeek is clear that this is a "Research prototype for reference only, not an official DeepSeek product." What to weigh: It's free and open, and the "one agent for the whole team" idea is worth studying if you juggle many agent chats. But it's research code that runs commands on your machine, setup is long (your own coding agent is meant to walk you through it), and DeepSeek's own article warns that a prompt injection, meaning hidden instructions planted in text the agent reads, could let an attacker take over connected machines. Try it only on a test repo.

2. OpenAI is handing out free "banked" Codex resets, and you have to claim them yourself. OpenAI's Tibo Sottiaux, who leads Codex, posted "Banked 1/2" at 5:43 PM PT Saturday, after saying on Thursday night that the "Next two days will be resets… banked…". Users are reporting the first one arrived, with one writing that it "landed on Halloween." OpenAI's help page explains that a banked reset is "a free, one-time reset saved to your account for later use." Getting one doesn't refresh anything; you choose when to use it before it expires. On the Plus plan, it refills both your 5-hour and weekly Codex allowances. On Pro plans, it refills the weekly one. It only works once at least one of those allowances is below 80%; otherwise you'll see "Your usage does not need a reset right now," and the reset stays saved. Using one can also move your next regular refresh time. It isn't cash or API credit. This follows our Oct 31 post on ChatGPT and Codex moving to shared usage limits in December. Try it: In the Codex desktop app, open Settings > Usage & billing (on the web, Usage settings) and look for "1 reset available." In the Codex CLI, type /usage and choose Redeem reset. The mobile app can't do it. Save it for a heavy day, and check the expiration date shown.

3. DeepSeek's web search and page-fetch tools now cost money, including when Claude Code or Codex use them. DeepSeek's API offers two tools for agents: Search, which returns web results from DeepSeek's own search engine, and Fetch, which turns a single web page into text. Each costs $0.75 per 1,000 successful requests, according to DeepSeek's docs. You're billed only when a request succeeds and returns something; failures, empty results and timeouts are free. The docs say that when you use the DeepSeek API through DeepSeek Harness, Claude Code or Codex, built-in search calls hit these tools automatically. Those calls were free, but "per-request billing starts at 00:00 Beijing time on November 1, 2026," which was 9 AM PT Saturday. What this means for you: If you've pointed Claude Code or Codex at DeepSeek's models to save money, web lookups now show up on your bill. At these prices it takes 1,000 searches to cost 75 cents, so it's small. Still, an agent that searches in a loop can rack up calls, so keep an eye on your usage page.

4. Chatbots lean on a small, AI-built voter guide two days before the U.S. election, The Washington Post found. In the Post's tests, done with the data firm Promptwatch, ChatGPT cited a site called the Ballot Brief in 22% of its answers about the election, and Microsoft's Copilot cited it in 75%. The site's founder used teams of AI agents to build profiles of thousands of candidates. The Post found mistakes on it: one senator was credited with supporting a bill that passed before he took office (an error Copilot repeated), and two different Robert Garcias running in California were merged into one profile. The bots were asked to make a voter guide for an address in each of the nation's 3,100 counties. They often put the address in the wrong congressional district or declined to answer. After the Post shared its results, ChatGPT's use of the site dropped to 4% in a second round of tests. Microsoft said it was looking into the results, while Anthropic's Claude started linking to the Ballot Brief more often in that second round. What to watch: Election Day is Tuesday, Nov 3. If you use a chatbot to research your ballot, check which site it's citing, and confirm your district and candidates with your county or state election office or your official sample ballot.

5. Google made Computer Use generally available in the Gemini API. Computer Use lets an AI agent operate a browser, phone or desktop app the way a person would. It looks at a screenshot, decides on a click or keystroke, and your own code carries out that action and sends back the next screenshot. Google's release notes say it moved to general availability (GA, meaning ready for production use) on Friday, Oct 30, across its Interactions and GenerateContent APIs. The docs name Gemini 3.8 Flash as the recommended model, with Gemini 4 Argon and the cheaper Gemini 3.5 Flash-Lite also supported. Each action comes with a short "intent" explaining why the model chose it. Built-in safety rules can block an action or ask a person to confirm it, for example before a payment or checkout. There's also an optional scan of screenshots for hidden instructions, the kind of prompt-injection trick mentioned in item 1. What this means for you: If you build automations for websites or apps (form filling, testing, data entry), Google now treats this as production-ready rather than a preview. Leave the confirmation step on for anything involving money or personal data.

6. A judge let the NAACP's lawsuit over pollution from xAI's data center move forward. A federal judge denied xAI's request to throw out the case, TechCrunch reports. The suit, filed for the NAACP by the Southern Environmental Law Center, says the SpaceX-owned AI company "installed and operated at least 35 combustion turbines and other sources of air pollution" to power its Colossus data center site outside Memphis "without ever obtaining the necessary preconstruction or operating air permits." xAI had argued that the gas turbines are mobile and so don't need permits, and that the NAACP had no standing to sue. Judge Debra M. Brown rejected those arguments, finding the turbines aren't "mobile sources" under Mississippi's air-quality plan, and wrote that "the plaintiffs' claims will proceed." A similar motion to dismiss from the Department of Justice was denied earlier this month. TechCrunch says it asked SpaceX for comment. What to watch: Whether the case ends up forcing xAI to get permits or cut back the turbines. It also fits a wider pushback against data centers' local costs, which TechCrunch notes could matter in Tuesday's midterm elections.

7. AI agents are now steering the Allen Telescope Array in the search for alien signals. Breakthrough Listen and the SETI Institute have brought AI agents into their daily search at the array of radio dishes in Northern California, Nvidia said on Friday. From one plain-English request, an agent can point the telescope's antennas, calibrate it, catch and fix errors in the incoming data, run AI models on the live stream and plot the results. Under the hood, a library from Nvidia's Holoscan platform pulls the data straight into graphics-card memory, so it's processed as it arrives instead of filling hard drives, at up to 2.4 terabits per second. The setup made what Nvidia calls the first real-time AI detection of a pulsar, a rapidly spinning dead star, in the Crab Nebula. It hasn't found aliens. When you'll feel it: Not directly, but it's a clear example of agents doing real lab work on live scientific instruments. The team is now testing open models, including Nvidia's Nemotron, for observatories that may someday run offline on the far side of the Moon.