On November 1, 2026, the DeepSeek Harness team posted a technical essay, "Exploring Super-agent Design: Towards a More Autonomous Agent System." A day later, Jen Zhu Scott quote-posted it with one line: "This is an obituary for Attention Economy."
That is a big claim for an engineering document. The essay never mentions ad markets, feeds or scrolling. It argues about agent architecture. But its core diagnosis, that human attention, not model capability, now caps how much work agents can do in parallel, pokes at the same scarce resource the ad-funded internet has monetized for two decades. To see why the line lands, and where it overreaches, start with the layer DeepSeek is redesigning: the harness.
What Harness Engineering Actually Is
An agent is not a model. The shorthand that spread in early 2026 is agent = model + harness, a formula the academic literature credits to LangChain's engineering writing. The model predicts tokens. The harness turns those predictions into multi-step action: the loop that keeps calling the model, the tools it may invoke, the context and memory it sees, the permission system that decides what runs freely and what waits for approval, the orchestration of sub-agents, and the extension points (skills, plugins, MCP servers, hooks) that let the system grow without a rewrite.
A July 2026 source-code study of eleven production coding agents, including Claude Code, Codex CLI, Gemini CLI, OpenHands, Aider and Hermes, read roughly four million lines of code and found the same seven subsystems in every one: agent loop, LLM integration, tools, memory and context, safety, orchestration and extensibility. None of them import a general-purpose agent framework, and the SKILL.md format now shows up in more of them (9 of 11) than MCP (8 of 11). The paper, Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents, closes with 18 design recommendations and a 90-line starter harness that implements ten of them. HackerNoon writers have been circling the same conclusion: agent reliability depends more on the harness than the model.
Developer NO1ennn read that paper and called harness engineering "the most f*cking DANGEROUS and accessible thing" he had ever seen. Both adjectives describe the same fact. Ninety lines buy a loop that can touch the filesystem, the terminal and the network. The model is rented or open-weight. The harness is the part you own, and the part that decides what the model may do when nobody is watching.
A Short History of the Layer That Matters
The scaffolding mattered before it had a name. ReAct-style reasoning loops and the 2024 SWE-agent paper showed that redesigning the agent-computer interface could move benchmark scores as much as swapping the model. Practitioners spent years building system prompts, tool schemas, permission policies and sandboxes without a shared term for the work.
The term arrived in early 2026. The Wavestone researchers behind the source-code study credit Mitchell Hashimoto with first using "harness engineering" in passing and LangChain's Vivek Trivedy with defining it. Hashimoto's version is practical: every time an agent makes a mistake, engineer a permanent fix into its environment so the mistake cannot recur. By mid-2026 the harness was an object of academic study, and DeepSeek pushed the idea to its limit with DeepSeek Harness, where even the agent loop is a plugin. There is no privileged core to patch.
NO1ennn's own long-form post shows what this looks like at night. He built the harness around his agent (context, tools, loop, evals, rules) but hated running his 412-eval suite every evening and triaging the failures. So he gave the night shift to an agent named Viktor: run the suite at 2 AM, compare failures with the night before, propose fixes as pull requests, and "do not merge or change anything on your own." Viktor prepares. A human signs. That one constraint separates a tool that runs when you run it from a system that owns a job while a person keeps signature authority.
What DeepSeek Actually Shipped
DeepSeek Harness (dsh) is an MIT-licensed, model-agnostic agent runtime whose README sums up its design as "Everything is a Plugin": the model adapter, tool registry, session log, sandbox, scheduler, UI and the agent loop itself. It runs on the Cordis plugin framework and launches as a local web UI with a single npx command. The repo now shows more than 270,000 GitHub stars.
The November essay, written by Yujia Qiao of the Harness team, names the single session as the next bottleneck. Models now finish complex tasks with less steering, so the constraint has moved. "A single task is limited by model capability, but multi-tasking is limited by human attention," the team wrote, adding that a person can actively track only three to five concurrent sessions. Past that point, more capable agents add context-switching costs instead of output.
The proposed Super-agent has four properties: continuous, productive, proactive and persistent. It is reachable across chat apps and GitHub comments without the user tracking sessions, it runs parallel tasks, it reacts to events without being prompted, and it recovers from failures. Architecturally, a long-lived Router session handles the human conversation and dispatches isolated Worker sessions for individual tasks. DeepSeek also contrasts its approach with Anthropic's "Decoupling the brain from the hands," noting that its design must first connect personal computers that are often only intermittently online.
The team released an MIT-licensed research prototype alongside the essay. In simulated workloads, the Super-agent finished tasks at a median of 12.4 minutes versus 20.4 minutes for a single-session agent. DeepSeek calls it "only a prototype for validating the technical approach, not an official DeepSeek product."
What the Attention Economy Actually Is
The idea traces to Herbert Simon's 1971 observation that a wealth of information creates a poverty of attention. For most of history, information was scarce. Once computers, networks and mass media reversed that, attention became the binding constraint.
In the 1990s, Michael Goldhaber's Wired essay "Attention Shoppers!" turned the observation into an economic claim: attention, more than information or even money, would become the real currency of cyberspace. Goldhaber described a pure attention economy of stars and fans. The commercial internet built something narrower: ad-funded platforms engineering notifications, recommendation feeds and infinite scroll to maximize time-on-site, then selling access to that time. HackerNoon contributors have argued that focus is becoming the class divide of this century for exactly that reason.
Has the Super-Agent Touched the Attention Economy Yet?
Almost not at all. The Super-agent is a prototype released days ago. Its benchmarks use simulated workloads, not real multi-user environments. DeepSeek Harness has serious developer traction as infrastructure, but it does not compete with consumer platforms or ad markets.
The idea is landing, though, while other arguments converge on the same weak point. IEEE Spectrum reports that digital advertisers may soon compete for AI agents' attention rather than humans'. Economists have modeled an "attention lemons" problem in two-sided ad markets once agents do the browsing. California Management Review describes a "trust economy" in which agents rewrite the rules of digital strategy. And on HackerNoon, one writer argues that AI coding already created a new attention market, while another warns that multitasking across coding agents drains the attention it was meant to save.
None of this dents Meta, Google, TikTok or the open-web ad stack this week. What changed is the recognition, visible in both DeepSeek's essay and NO1ennn's night shift, that the scarce resource is no longer only the hours a person spends scrolling. It is also the hours a person spends supervising the agents that were supposed to free those hours.
The Danger and the Accessibility
The same 90-line starter that makes harness engineering accessible makes it dangerous. A model without a harness is a chat window. A model with a good harness can read private files, run shell commands, open pull requests, send messages, schedule recurring work and recover from its own failures. The Super-agent hides the session-management tax so one person can keep more of that work in flight. Open, trivial-to-start harnesses plus architectures that demand less supervision lower the barrier to deploying systems with persistence and reach.
Viktor's "do not merge or change anything on your own" is one careful answer. It is not a prompt trick but a hard boundary in the harness: every action that matters waits for a human yes. Today the person who built the harness enforces that boundary. As Super-agent-style systems grow more continuous and proactive, the open question is whether the boundary stays a deliberate engineering choice or erodes under the pressure of "just let it finish."
Accessibility also means failure modes are no longer confined to well-funded labs. Anyone who can write a SKILL.md, configure a plugin or fork a 90-line loop can hand a model real-world effectors. Safety is one of the seven canonical subsystems precisely because the other six can do damage. When the harness is the product, its bugs are capability bugs: a context window that silently drops a constraint, a tool description that is slightly too broad, a permission policy that fails open. HackerNoon's own chronicle of the period, The Age of the Lobster, catalogs how quickly those effectors spread.
Will It Reshape the Attention Economy?
The structural question is more interesting. Today's agent interfaces keep the session as the unit of work: open a chat, assign a task, watch it, switch contexts, restate goals. The human is the router. The Super-agent moves routing into the system so the person works at the level of decisions and approvals. If that pattern spreads, three things follow.
First, attention per unit of work falls. DeepSeek's team says plainly that the bottleneck to further productivity is shifting toward human attention. A continuous agent that dispatches parallel work and surfaces results only when judgment is needed attacks the assumption that more capable agents must consume proportionally more human attention.
Second, the main consumer of interfaces may change. When agents browse, compare, buy and coordinate for us, value flows to interfaces that are machine-readable and criteria-satisfying rather than sticky. Content built for emotional engagement or variable-reward notifications matters less if the agent never scrolls. Media companies will face pressure to publish material that agents can parse, trust and act on.
Third, attention gets reallocated, not abolished. Goldhaber's point was that people's appetite for attention has no upper bound. Freeing attention from low-level monitoring raises the value of what remains for high-stakes judgment, relationships and creative work. Others warn of an "intention economy", in which platforms move upstream from capturing attention to forecasting and selling intentions before users fully form them.
The counter-forces are real. Ad and recommendation systems will adapt. Platforms can still monetize the moments when humans review agent output, set goals or resolve ambiguity. Synthetic agents that mimic human engagement can pollute metrics. Coordination between agents remains harder than individual competence. The harness layer simply moves the scarce resource one level up: from the content that captures attention to the systems that decide what agents attend to on our behalf.
The Verdict: Calling the Super-agent essay an obituary is premature.
Platforms engineering for maximum human engagement still dominate. The prototype has not proven it scales, its tests are synthetic, and the ad infrastructure is vast. Obituaries get written after the fact. This is a design document.
It is, however, a precise diagnosis of a moving bottleneck and a concrete architectural response. "Multi-tasking is limited by human attention" names the same scarcity Simon described in 1971, now appearing inside the agent stack. If systems that route work, act proactively and need less supervision become the default interface to the digital world, the logic that rewards capturing human attention for its own sake loses force. Value moves toward systems that satisfy agent criteria, earn trust for delegated action and give attention back.
Harness engineering makes that shift more accessible and more dangerous at once. Open-source runtimes lower the cost of giving a model real effectors. The Super-agent lowers the ongoing attention cost of keeping them running. Together they move the scarce resource from "who can capture human eyes" to "who designs the boundaries that still require a human yes." NO1ennn kept the signature. Whether that stays the default or becomes a setting most people switch off is the question the next few years of harness design will answer.
