DeepSeek's Super-agent: Another R1-Style Shock?

On Saturday Night, DeepSeek Didn’t Drop a Model

Whenever DeepSeek puts something out, the AI industry tends to shudder a little. This time, what it put out wasn’t a model — it was an agent. On the evening of Saturday, October 31, DeepSeek’s Harness¹ team posted a long piece on WeChat titled “Exploring Super-agent Design: Toward More Autonomous Agent Systems,” and on the same day released a research prototype called dsh-superagent-prototype on GitHub under the MIT license. Bloomberg reported it as DeepSeek’s answer to Anthropic’s Claude Code — an agent that watches developer group chats, and the moment it spots feedback about a coding problem, offers a fix without anyone asking.

Here’s some background. Around May this year, DeepSeek formed a Harness team that used Claude Code as its internal benchmark, and on August 13 it released the agent runtime deepseek-harness under the MIT license. It also raised money. According to Bloomberg and Reuters, DeepSeek raised roughly ¥100 billion (about $15 billion) in its latest funding round, with CATL and Tencent as the major backers. That’s twice the size of its first external round (about ¥50 billion) closed in June, and the company is also preparing for an IPO.

I want to ask just one question today. Can this Super-agent deliver a shock comparable to the one DeepSeek-R1 caused in January 2025? To answer that, I need to check two things first: what Super-agent actually is, and exactly where R1’s shock came from.

What the Super-agent Is Really Trying to Save: Human Attention

The design document doesn’t start from model performance — it starts from human attention. The harness team figures a single person can realistically track 3–5 active sessions at once. The ceiling on any one task is set by the model, but the ceiling on how many tasks you can run is set by human attention. That’s why the stated principle is “take as little human attention as possible, and give as much value as possible.” The idea is to take up as little of a person’s attention as it can while giving back as much value as it can.

The architecture has two layers. A standing Router talks with the human and hands out work; isolated Workers, running in parallel, actually do the tasks. The Router never breaks a job into steps — it only decides whose request goes to which Worker. That’s why the design document recommends a lightweight model for the Router and the strongest available model for the Workers.

Inputs aren’t prompts either — they’re events. Chat messages, code-review comments, CI² results all come in as triggers; code filters them first, and only what survives goes to the model for judgment. A separate Watchdog session periodically checks whether a Worker that reports “in progress” is actually making progress. Agent state lives in a cloud database while commands are executed by a small daemon on the user’s own machine — splitting the two so the agent stays online even when the user’s device is powered off.

There’s a number worth noting, too. In a synthetic test that mixed roughly 20 genuine task requests into about 300 messages from group chats and GitHub CI, the Super-agent finished 95% of requests within 15.7 minutes. A single session equipped with sub-agent tools took 29.5 minutes (both figures are p95³). That said, the test swapped out actual coding for a fake tool that simply burns time — what it measures is the ability to split and coordinate work, not code quality. The harness team itself called this a preliminary reference figure.

The “it moves without being asked” detail that Bloomberg highlighted deserves to be read with some nuance. The design document does describe a case where an internal DeepSeek instance, watching a public group chat, tracked down the cause of a bug nobody had reported and even opened a fix PR unprompted. The publicly released prototype, however, is far more cautious. According to its README, this agent acts only when a human has requested it or when rules the team has written down permit it, and it speaks up only when someone needs to act or when there’s a result worth writing. Its supported scope is also narrow — a single GitHub repository, plus an optional Feishu integration. DeepSeek itself has said this is research code, not an official product; it differs from the internal version, and parts of the design document were never implemented. The remaining work items it flagged are scaling the Router and security — and it noted, in its own words, that if the cloud-side loop were ever breached, every connected device would be at risk.

What R1 Actually Shook Wasn’t Model Performance

There’s a detail many people miss. What shocked the industry was R1 — not DeepSeek’s general-purpose model. And even R1’s shock didn’t come from its performance per se. It was the development cost, the decision to release it as open source, and the serving cost that landed the blow.

On January 20, 2025, DeepSeek released R1, a reasoning model said to rival OpenAI’s o1, and published its weights and technical report under an MIT license — explicitly inviting distillation⁴ and commercial use without restriction. The cost story, though, actually came from the technical report of V3, the base model released a month earlier. It stated that the full training run consumed 2,788,000 H800 GPU-hours. At $2 per hour, that works out to roughly $5.58 million — which is how the figure “a frontier model built for $6 million” spread through the market. Strictly speaking, that number covered only the training compute, not the total development cost including research and experimentation. Still, the message the market took away was unambiguous: this could be done for far less money than anyone assumed.

Serving prices pointed the same direction. The R1 API charged $0.55 per million input tokens and $2.19 per million output tokens — nearly 30 times cheaper than o1 on output pricing. On January 27, Nvidia’s stock dropped nearly 17%, wiping out about $589 billion in market cap in a single day.

If you had to name the moments that genuinely shocked LLM developers, the list would be short: Google’s Transformer in 2017, BERT in 2018, Meta’s release of Llama 2 under a commercial-use license in 2023, and the release of DeepSeek-R1. DeepSeek’s place on that list comes from having delivered a reasoning model that could be grafted onto anything. The six distilled models released alongside R1 — built on Alibaba’s Qwen and Meta’s Llama, ranging from 1.5 billion to 70 billion parameters — included a 32B and a 70B version that DeepSeek claimed performed on par with o1-mini. In other words, reasoning ability was no longer locked inside one company’s model; it had become a technique that could be transplanted into someone else’s.

Get the Guardrails Right, and Price Is All That’s Left

Now let’s look at where things stand today. Something has shifted as models approach superintelligence and work can be delegated through instructions alone. Call it a harness or call it a skill — once you hand a capable enough model clear guardrails written in Markdown, performance above a certain threshold is basically assured. Which means the only variable left is price. I think that’s exactly why Chinese models like Kimi, GLM, and DeepSeek are so popular with American companies, even amid all the security-threat controversy. They’re tens of times cheaper.

I checked that “tens of times” against the official price sheets. These are dollars per million tokens, as of November 1.

Model	Input	Cache Read	Output
Claude Opus 5.5	4.00	0.20	20.00
Claude Haiku 5.5 (prompt ≤100K tokens)	0.10	0.01	0.50
Claude Haiku 5.5 (prompt >100K tokens)	0.50	0.05	2.50
DeepSeek V4-Pro (peak / off-peak)	1.32 / 0.66	0.044 / 0.022	3.96 / 1.98
DeepSeek V4.1-Flash (peak / off-peak)	0.30 / 0.15	0.006 / 0.003	1.20 / 0.60

Against Opus 5.5 — the model usually tasked with the hardest coding work — DeepSeek V4.1-Flash is 27 times cheaper on input and 33 times cheaper on output, at off-peak rates. For cache reads, which agents rely on when they repeatedly re-read the same context, the gap widens to 67 times. The structure means the gap grows largest not in a single one-off chat exchange, but in agentic work that rereads the same repository dozens of times.

Two caveats apply, though. One is DeepSeek’s peak window: weekdays 01:00–04:00 and 06:00–10:00 UTC, which translates to 10 a.m.–1 p.m. and 3–7 p.m. KST — almost exactly Korean business hours. Prices double during that window, shrinking the gap with Opus to 13–17 times. The other is Anthropic’s Haiku 5.5: if the prompt stays under 100K tokens, both its input and output rates beat Flash. Past 100K tokens, Flash becomes cheaper again — and coding agents that ingest an entire repository routinely cross that line. So the “tens of times” gap holds between frontier models and Chinese models specifically; it doesn’t describe American models as a whole.

The market has followed this math. According to a CNBC report in July analyzing OpenRouter⁵ data, the share of tokens that American companies routed through OpenRouter to Chinese models exceeded 30% every week after February 8, peaking at 46%. The 12-month average before that had been 11%. OpenRouter attributed this to Chinese open-source models being 60–90% cheaper than Anthropic’s and OpenAI’s flagship models. Washington is looking at this from the opposite direction: in July, the Treasury Secretary raised the possibility of sanctions on Chinese models, and in September, reports surfaced that House committees were investigating the spread of Chinese AI usage.

OpenRouter is a broker where developers swap models in and out, so it doesn’t represent the entire market, and token share isn’t the same as revenue share. Still, it offers a fairly honest window into what gets chosen when switching costs are low.

DeepSeek understands this dynamic and is acting on it. DeepSeek’s API price sheet lists Claude Code and Codex right alongside its own harness as pathways for calling its search and fetch functions — and it announced that calls through this path, free until now, would start being billed per call from November 1. In other words, DeepSeek has no objection to its own model running inside a competitor’s harness. Its Super-agent prototype shows the same flexibility: it defaults to whatever model is configured in the harness, but lets you choose separate models for the worker and the evaluator.

Oswarld’s Lens

I’m actually working by this math in practice. As of September, tasks that run long and repetitive work—things I used to hand off to GPT’s Luna-tier models or Sonnet—now go to DeepSeek and GLM. Overall planning, orchestration, and initial blueprints still go to Claude or GPT-6. Even burning trillions of tokens on Ollama Cloud costs overwhelmingly less than Claude or GPT-6, and if you set up your harness files—things like Skill.md or soul.md—properly, performance holds up just fine.

There’s something I always say: not everyone, and not every task, needs a superintelligence. There’s a model suited to the job and the environment. Even DeepSeek’s own design writeups recommend putting a light model at the router and your strongest model at the worker. In other words: deploy different grades of intelligence for different roles.

So could a Super-agent become the next R1? I don’t think it’ll be the same kind of shock. With R1, the development cost, the scope of the open release, and the serving price were all numbers the market hadn’t priced in. This time, there’s no number that surprises anyone. Companies that have open-sourced coding harnesses already exist—OpenAI’s Codex CLI among them—everyone already knows DeepSeek is cheap, and Super-agent hasn’t disclosed a single line about its costs.

What’s new isn’t the numbers—it’s the unit of measurement. Alongside how many minutes an agent worked, people have started measuring how many times a human had to step in and check. That kind of shift is less likely to move stock prices in a day and more likely to quietly seep into other companies’ product design documents.

Closing

When R1 came out, the question people asked was “Wait, this works at that price?” The question a super-agent raises is a little different. It’s how many times I have to check in on this work. Still, this prototype is only research code handling a single repository, and even DeepSeek hasn’t answered how to safeguard the authority to execute commands in place of a human. Before you bring an agent that works without being asked onto your team, you need to see that answer before you see the price tag.

---

💬 Reader, which model are you pairing with your coding agent these days? If you’ve shifted any work over to a Chinese model, tell us what kind of task it was in the comments.

📨 If a colleague of yours is weighing the cost of adopting an agent, pass this piece along.

---

References & Further Reading

Primary sources

• DeepSeek, “dsh-superagent-prototype README”, GitHub, 2026.10.31. ··· This is where the prototype’s scope, its “research code only” caveat, and the stated principle of conserving human attention all appear.
• DeepSeek Harness Team, “Exploring Super-Agent Design: Toward More Autonomous Agent Systems”, WeChat, 2026.10.31. ··· The original design document, in Chinese. It contains the router-worker architecture and the synthetic benchmark figures.
• CellCog, “DeepSeek’s Super-agent: One Router, Many Workers”, 2026.11.1. ··· An English rendering of the design piece’s structural tables and test conditions. Start here if the Chinese original feels daunting.
• DeepSeek, “Models & Pricing”, accessed 2026.11.1. ··· Covers peak/off-peak pricing and guidance on calling the model via Claude Code and Codex.
• Anthropic, “Pricing”, accessed 2026.11.1. ··· Token pricing for Opus 5.5 and Haiku 5.5.
• DeepSeek, “DeepSeek-R1 Release”, 2025.1.20. ··· The original announcement of the MIT license and the six distilled models.
• DeepSeek-AI, “DeepSeek-V3 Technical Report”, arXiv, 2024.12. ··· The source for the figure of 2,788,000 H800 GPU hours.

Background

• MIT CSAIL Alliances, “DeepSeek: What You Need to Know”, 2025.1.28. ··· Summarizes the drop in Nvidia’s market cap immediately after R1.
• codecentric, “Open Source hits Billion-Dollar Market: DeepSeek-R1 is shaking up the AI world”, 2025.1.29. ··· Compares the output pricing gap between R1 and o1.
• AI Weekly, “Chinese Models Take 30%+ of US OpenRouter Token Use Since Feb 8”, 2026.7.7. ··· Summarizes CNBC’s analysis of OpenRouter data, with the caveat that it reflects a single platform’s numbers.
• Let’s Data Science, “DeepSeek reportedly raises $15 billion in round backed by CATL, Tencent”, 2026.10.22. ··· Rounds up Bloomberg and Reuters reporting on the ¥100 billion funding round.

Related past issues worth reading

• What on Earth Is a Harness?! On the Harness That’s Supposedly Boosting AI Productivity ··· The issue that first laid out today’s premise of “thin harness, thick skills.”
• If It’s Not in the Group Chat, It’s Not a Product ··· Frames the bottleneck in agent adoption as a verification problem — connects directly to why Super-agent is measuring human attention.
• The Legal AI OpenAI Invested In Is Secretly a Chinese Model ··· A concrete case of a US company defaulting to a Chinese open-weight model.

---

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. Harness: The software scaffolding that wraps a language model to make it act as an agent. It runs the model repeatedly, handles files and tools, and manages context and safeguards. The name comes from the harness put on a horse. ↩

2. CI (Continuous Integration): A system that automatically builds and tests code every time a developer pushes a change. A failure triggers an immediate alert — which, for an agent, is a signal that there’s work to do. ↩

3. p95: The value at the 95th percentile when all requests are sorted by how long they took. It means “95 out of 100 requests finished within this time,” capturing the slow tail that averages hide. ↩

4. Distillation: A technique that trains a smaller model using a larger model’s outputs as its textbook. The DeepSeek-R1 release shipped alongside smaller models from other companies into which this method had transplanted its reasoning ability. ↩

5. OpenRouter: An intermediary platform that lets users access AI models from multiple companies through a single API. Because it publicly reports how much each model is used, it’s frequently cited when tracking model usage trends. ↩