SI Agent News

• Arken, a very early open-source project: a cryptographic identity layer for AI agents calling MCP servers (the connectors agents use to reach tools) and APIs. Usable now, but it had just 1 point and no comments on Hacker News when we saw it. Status: REPORTED, and early. Source: GitHub.
• TMTPost commentary: argues the ceiling for AI agents is no longer compute but ecosystem walls, meaning apps that will not let agents in. Opinion, not news. Source: tmtpost.com.
• Tech Traveler asks whether AI that solves problems really understands them. A reflective piece with nothing to act on this week. Source: techwalker.com.

DeepSeek has open-sourced a coding agent that fixes bugs unprompted. DeepSeek (a Chinese AI lab) put out Super-Agent (an open-source coding agent). technews.tw reports that its Harness team published the code on its WeChat official account on October 31, and that the agent fixes bugs proactively, without being prompted. The outlet frames it as a direct challenge to Claude Code (Anthropic's coding agent). Status: REPORTED. We have a news report, not DeepSeek's documentation. The model behind it, the license and the hardware needs are STILL OPEN. Open-source code is not the same as open weights, so read the license file before this goes near client work. Price not stated. It is worth it for developers who can review a code change; if you cannot read a diff, wait. If you do try it, use a throwaway repository first, and make the agent propose changes for a human to approve instead of committing directly. Source: technews.tw.

What this means for you:

• Small business owner: ask your developer whether they are testing proactive bug fixing, and who reviews every change.
• Job seeker or student: an open-source agent's code is free reading on how these tools are built, and a strong interview talking point.
• Creator: it could cover overnight breakages on a site or plugin, but keep it away from anything touching payments.

The money behind it. 36Kr Global's weekly roundup reports DeepSeek closed an A+ round of over 100 billion yuan, a number with eleven zeros. The same headline also lists Geely Auto's vehicle exports of 798,000, though the headline is cut off in our copy, so we do not say more about that. Status: REPORTED, from a weekly roundup with no company statement seen. If true, it suggests investors think agents, not just models, are where the next money goes. For ordinary users the practical meaning is continuity: a well-funded lab is more likely to keep maintaining a tool you start to depend on. Funding says nothing about price, license or quality, so do not plan around the figure, and do not treat it as proof that Super-Agent is good. Source: 36Kr Global.

Someone actually tried it. Sina Finance ran a hands-on piece, translated 'I'm actually raising a DeepSeek super agent,' with a step-by-step tutorial. The author says it works automatically without being @-mentioned, meaning you do not tag it to get its attention. That matters because most agents today wait for a command, and one that acts on its own changes the question from 'can it do the task?' to 'when should it be allowed to act?' We have not run it and do not have the full tutorial, so we do not repeat the steps. One person's experience is a demo, not a benchmark. If you follow the tutorial yourself, check each step against official DeepSeek material and note what permissions you hand over. Status: REPORTED. Source: Sina Finance.

Agents are starting to look alike. A piece on 80aj, titled 'Five companies' Agents are becoming more alike: DeepSeek is the fifth,' argues that agent products from different companies are converging, with DeepSeek the latest to join the pattern. We have the headline and framing, not the full text, so we cannot say which four other companies it names or what features it compares. Status: REPORTED, and it is one outlet's analysis. If the claim holds, it is useful for buyers: when the products behave similarly, the real differences shrink to price, license, data handling and how easily a human can review and undo what the agent does. Keep your prompts, project notes and workflows in a form you can move, so you can switch tools without starting over. Source: 80aj.

Figma opens its canvas to coding agents. Startup Fortune reports that Figma (a design tool) is opening its canvas to AI coding agents, answering rival Canva's pivot. The direction is real: design files becoming something an agent can read and build from, instead of a picture a developer re-types into code. For a small business or creator, the likely payoff is a shorter path from a mockup to a working page, with less handoff between designer and developer. The report does not say which agents are supported or what it costs. For open-model users the concrete question is whether an agent you run yourself can plug in, or whether access is limited to a few approved partners. Status: REPORTED. Source: Startup Fortune.

Can your agent be switched off? The Strategist, from the Australian Strategic Policy Institute (ASPI), argues that organisations need to prove their AI agents can shut down. We have the headline and argument, not the full piece, so we do not attribute specific recommendations to it. The point is directly relevant to this week's news. An agent that acts without being asked, like the one DeepSeek describes, is only safe to run if someone can stop it quickly and know what state it left behind. Before you deploy any agent, ask four plain questions. Who holds the off switch? Have you tested it? What happens to a half-finished task? Can you see a log of what it did? If the vendor cannot answer, treat that as your answer. Status: REPORTED, an opinion and analysis piece rather than a finding. Source: The Strategist (ASPI).

Giving agents an identity. A Show HN post on GitHub introduces Arken, described as cryptographic identity for AI agents calling MCP servers and APIs. MCP (Model Context Protocol) is a common way for agents to connect to outside tools and data. The idea behind the title is that a service can verify which agent is calling it, instead of trusting a shared password. We have not tested it or reviewed the code, and we do not describe how it works beyond that summary. Status: REPORTED, an early community project, not an audited product. Why it matters: once agents act on their own, you need to know which one did what, and revoke one without breaking the rest. Developers building agent workflows should read the repository's documentation and treat it as an experiment. Everyone else can simply note that this problem is now getting attention. Source: GitHub (Show HN).

Is compute still the limit? TMTPost asks whether compute is no longer the ceiling for AI agents, and whether ecosystem walls are the real constraint. The headline poses a question, and we have not seen the full argument, so read it as one outlet's framing. The practical reading is plausible, though. An agent is only as useful as the apps, files and services it is allowed to reach, and closed platforms can limit that no matter how strong the model is. That connects to the Figma story above: the news there is access, not intelligence. When you evaluate an agent, check which tools it can connect to, whether those connections need the vendor's approval, and whether your data can leave. Status: REPORTED, analysis rather than a measurement. Source: TMTPost.

Solving is not understanding. Tech Traveler asks whether AI can solve problems but really understand them. The headline signals a familiar debate, and we have not seen the full piece, so we do not claim what evidence it uses. The useful takeaway for daily work is cautious. An agent that fixes a bug may still have fixed the symptom, not the cause, and it will not always tell you the difference. Check its work: ask it to explain the change in plain words, run the tests yourself, and try a slightly altered version of the problem to see if the fix still holds. That habit costs minutes and is the best protection against confident wrong answers, especially with proactive agents that change things before you look. Status: REPORTED, an opinion piece. Source: Tech Traveler.

Prove your agent can stop. The Strategist, the analysis site of the Australian Strategic Policy Institute, argues that organisations need to prove their AI agents can shut down. Any organisation considering agents, especially in national security or critical infrastructure, should give them no system access until sure they will obey an order to stop. Status: CONFIRMED that the article makes this argument. Whether any given agent reliably obeys a stop order is STILL OPEN. Source: The Strategist.

Why it bites harder now. A chatbot acts only when you speak. A proactive agent has its own trigger, so it may already be acting when you decide you want it to stop. That is the link to today's lead.

What proof could look like (our analysis, not ASPI's checklist):

• Stop it mid-task and confirm it really halts.
• Stop it after a long chain of steps, where agents tend to drift.
• Check what it leaves behind: half-applied changes, open connections, credentials that still work. A halted agent with a live key is not halted.

Self-hosting cuts both ways. If you run an open model yourself, you own the off switch: kill the process, revoke the key, no vendor in the loop. But you also have no vendor safety layer. Nobody built that off switch for you.

Who gains, who loses. Security and risk teams gain, as does anyone burned by an overeager automation. Teams chasing speed lose a few days. That is a fair trade for anything touching real systems, less so for a hobby repo. Takeaway: treat stop as something you test, not something you assume.

For builders: what is inside a proactive coding agent. 80aj.com argues DeepSeek is the fifth company whose agent design is converging on the same shape (80aj.com). Our summary does not spell out that shape, so what follows is general analysis of how such agents are typically built, not DeepSeek's documented design.

The loop. An agent is a model in a cycle: look at the situation, decide, use a tool, check the result, repeat. The harness is everything around the model: tools, memory, permissions, stop rules. DeepSeek's team is named Harness, which fits.

What proactive adds. Normal agents start when a human sends a message. A proactive one starts from an event: a test fails, an error hits the logs, an issue is filed. That is a trigger layer, simple code watching for events. Then a filter decides whether the event is worth acting on; the loop gathers context, writes a patch, runs the tests, and ideally returns a pull request (a proposed change) instead of editing code directly.

Why five companies converge. Same constraints. Models are decent at calling tools. Bugs have a built-in referee, the test suite. And the risk pushes everyone toward a human approval step.

Novel or incremental? Mostly incremental. The loop is not new. What is new is dropping the human nudge and, if the report holds, releasing the harness as open source so you can read and change the trigger and permission code.

Limits: false alarms (fixing what was not broken); cost, since a watcher that wakes the model constantly keeps spending, tokens if you pay per use, GPU time and electricity if you self-host; and looping on the same failing fix.

Outlook (analysis, not fact): competition shifts from whose model is smartest to whose harness is safest and cheapest to run, which favors people who can inspect the code. STILL OPEN: whether DeepSeek's version delivers that; we have only seen reports.

Lesson, not news: a safe way to try a proactive bug-fixing agent. This is an illustrative setup, not tested by us and not DeepSeek's configuration.

For builders: a developer or small team with one repo they can afford to experiment on.

• 1. Pick a harmless repo. Nothing in production, nothing with customer data.
• 2. Give the agent its own identity. A separate account or token with least privilege: it can open branches and pull requests, it cannot merge, it cannot see secrets.
• 3. One trigger only. A failing test run on the main branch, not every event.
• 4. Comment-only week. The agent posts its proposed patch as a comment and never touches a branch.
• 5. Caps. Two attempts per failure, a time limit per run, a daily ceiling on runs.
• 6. Stop drill. Kill it mid-task, confirm it halts, revoke its token, confirm it is locked out.
• 7. Log everything. At week's end count patches you would accept, patches you would reject, and false alarms.

agent:
  trigger: test_failure_on_main
  permissions: [create_branch, open_pull_request]
  denied: [merge, read_secrets, edit_config]
  mode: comment_only   # switch after one week
  max_attempts_per_failure: 2
  max_runtime_minutes: 10
  max_runs_per_day: 5
  log: all_actions

Expected output: a short report and a diff per failure. How to verify: the stop drill passes every time, and the log shows no action outside what you allowed. If accepted patches are not clearly more than rejected ones, do not let it open pull requests. Limits: it only suits bugs that tests can catch, and one week is a small sample. Cost: price not stated for DeepSeek's agent; the software may be free but the model behind it may not be.

Don't wake the model at all. The cheapest saving is not a smaller model; it is a plain, non-AI filter in front of the agent. A simple rule checks: is this a real failing test on main, and is it new? Only then does the model get called. That filter decides how often you pay, in tokens or GPU time.

One number (our estimate, not from a source): if triaging one failing test by hand takes about ten minutes, five a week is about fifty minutes. An agent that drafts a first pass you only review could win back part of that. Time one week of your own triage first; your numbers will differ.

Options from different companies: DeepSeek's open-source Super-Agent if the reports hold, Anthropic's Claude Code, or OpenAI's Codex. Prices are not in today's stories, so check each pricing page.

What you give up: you move effort from writing to checking, and a bad patch you approve costs more than the ten minutes you saved.

Prompt clinic. The weak version: Fix the bugs in my repo. No scope, no limits, no definition of done, and dangerous for a proactive agent.

The better version (illustrative, not tested):

You are a cautious maintenance assistant for this repository.
Trigger: the test named in the failure report below is failing on the main branch.
Task: find the smallest change that makes that one test pass without breaking others.
Rules: do not edit files outside the src folder. Do not touch configuration, secrets or dependencies. Make at most two attempts, then stop and write a short report.
Output: a proposed patch as a diff, a two-sentence explanation of the cause, and the exact test command you ran with its result.
If you are not sure, say so and propose nothing.

Why it works: scope (one test), boundary (one folder), budget (two tries), proof (the command and result), and permission to do nothing, which cuts false fixes. Customize: swap the folder rule for your layout. Quality check: re-run the test command yourself; do not trust the report.

No new paper in today's stories, so this is a short lesson on reading claims, not news. Today's claim is that an agent fixes bugs without being prompted. Reports like that rarely say how often it is right. A benchmark (a standard set of test tasks used to compare tools) tells you about its tasks, not yours.

Decision this week: take ten bugs you have already fixed, run the agent on them, and count how many it fixes, how many it breaks, and how many it ignores. Ten is a small sample: enough to catch an obvious mismatch, not to prove quality. My estimate is roughly an afternoon's work.

The direction is clear: agents are moving from answering you to acting on their own, and designs are converging. The counterweight is governance, the case for proving an agent will stop.

Watch next, none confirmed yet (STILL OPEN): whether DeepSeek's repository shows a clear license and setup documentation; whether the reported funding round gets a company confirmation; and which agents Figma actually supports.

Listening every day is learning with a reason: today it is a narrow trigger to cut cost and a stop test to keep control, drawn from reports in China and Australia.

My agent fixed a bug without being asked. Now I'm asking it to stop fixing things.

That's it for today. Tomorrow we're watching whether DeepSeek's repository publishes a clear license and setup docs, and whether the funding report gets confirmed. If this made you smarter, share it with one person who needs it. Thank you for spending part of your day with us.

SI Agent News — a podcast newsletter. Made by AI.