The missing
protocol — and a
call to build it.

Every serious conversation about AI and work arrives at the same impasse. The economists project displacement. The politicians debate redistribution. The AI companies talk about augmentation. And then everyone goes back to what they were doing, because the gap between the macro-level debate and the micro-level reality — a specific person whose specific job is being automated, right now, in a specific company — is enormous and nobody has a concrete bridge across it.

The policies being discussed — universal basic income, AI taxation, mandatory retraining — operate at the level of nation-states and decades. The displacement is happening at the level of workflows and quarters. The mismatch in scale is one of the most dangerous aspects of the current moment. By the time the policy response arrives, the disruption will already have reshaped the labour market in ways that are very difficult to reverse.

What is missing is not a policy. It is a protocol. A technical standard, like OAuth for authentication or SMTP for email, that creates a defined interface between AI systems and human workers — one that makes it structurally easy for any AI agent to route work to a human when that is the better choice, and to pay that human fairly for doing it.

This is not a new idea at the concept level. Amazon Mechanical Turk demonstrated in 2005 that human microtask markets could work at scale. Every content moderation system at every major platform routes certain decisions to human reviewers because AI cannot handle them reliably. The concept is proven. What does not exist is the standard — the open protocol that any agentic AI system can call, regardless of which company built it, to connect a task to a human worker who can complete it.

What task-bridge is trying to be

The task-bridge repository on GitHub is an early attempt at exactly this. The core idea is simple: a standardised JSON interface that any AI agent can call to declare that a task has exceeded its confident capability threshold and route it — with context, with compensation parameters, with a structured description of what is needed — to a human worker pool.

Think of it as a structured handoff layer sitting between AI execution and human execution. The agent does not just fail and stop. It does not hallucinate an answer it is not confident in. It calls the protocol, describes the task it cannot complete reliably, specifies the skills needed to complete it, and routes it to a marketplace of human workers who have those skills — paying them fairly from the economics that the AI operation was already budgeting.

The elegance of this model is that it does not require AI companies to limit their systems or governments to pass new laws. It requires AI developers to build a confidence_threshold mechanism into their agents — something many already have in some form — and connect that threshold to a standardised routing protocol rather than a failure state.

Critically, the flow is bidirectional. Task-bridge is not only an outbound channel for AI systems to offload work they cannot handle. It is equally an inbound channel for humans who need work completed with a human touch — tasks that require physical presence, local knowledge, cultural nuance, or the kind of judgment that only comes from lived experience. A person who needs their elderly parent's medication collected from a pharmacy on the other side of the city, a business that needs a document reviewed in a language its AI tools do not cover reliably, a contractor who needs a physical site inspection — all of these can post to the same board. The system does not distinguish between AI-originated requests and human-originated ones. It distinguishes between tasks and the qualifications needed to complete them.

The approval layer sits in the middle of both flows. Before a task hits the board — whether it came from an AI agent that hit its confidence threshold or from a human who knows they need another human — it passes through a structured validation step. Is the task description complete enough for a worker to act on? Is the compensation appropriate for the work and location? Is there a clear completion criterion so both sides know when the job is done? Only tasks that pass this validation reach the board. This is not bureaucracy for its own sake. It is the mechanism that keeps the board signal rather than noise — and that ensures workers are not accepting tasks that will waste their time or underpay them.

Once on the board, tasks are visible only to workers whose verified skills match the requirements. Completion triggers a timer. If the timer lapses without completion, the task returns to the board — available to the next qualified worker, or the original worker can request a reset with a revised timeline. The payout calculation accounts for the full cost of completion: time, travel, materials, platform fee. Not a platform-optimised minimum. A genuinely fair total that reflects what it actually costs a real person to do real work.

The human worker side of the equation is equally important. A protocol that routes tasks to humans only works if there is a structured marketplace on the other side — one where workers have verified skills, transparent compensation, and genuine agency over which tasks they accept. This is not Amazon Mechanical Turk, where workers are anonymous and compensation is often degrading. It is a skilled labour exchange where the skills are real, the compensation is fair, and the workers are partners in an AI pipeline rather than a residual category.

The ecosystem that is already emerging

Task-bridge is not the only project approaching this problem. A small but serious ecosystem of human-in-the-loop protocols is already building, and the overlap is meaningful.

HITL Protocol by rotorstar frames the problem from the service side: "HITL Protocol is to human decisions what OAuth is to authentication — an open standard connecting Services, Agents, and Humans." It proposes a standardised flow for services to request human input when an AI agent is about to make a consequential decision — cancel an insurance policy, book a non-refundable trip, wire money. The human stays in their preferred messaging channel — Telegram, WhatsApp, Slack — and taps a decision button there. No wall of text. No blind agent decisions.

AgentRQ approaches it from the task management side: a modern, high-performance platform designed for seamless collaboration between human operators and AI agents, where humans can break down complex goals into manageable tasks and delegate work directly to AI agents — or, critically, for AI agents to surface tasks back to humans when they need review or approval.

Every agent framework already has its own mechanism for gating what the agent itself is about to do — OpenAI's Agents SDK can pause a tool call, LangGraph has interrupt(), OpenClaw asks before running a dangerous command. What none of them have is a standardised interface for routing work to humans outside the framework. That is the gap task-bridge is aimed at.

Why this needs to be open

The most important word in the task-bridge proposal is open. Not open-source in the sense of a GitHub licence, though that matters. Open in the sense that no single company controls it.

If Anthropic builds a human handoff system, it will route tasks to Anthropic-preferred workers on Anthropic-preferred terms. If OpenAI builds one, same story. If a startup builds one with VC backing, the economics of the marketplace will be shaped by investor return requirements rather than worker welfare. Every closed implementation of this idea has the same failure mode: the platform captures the value that should flow to the workers.

An open protocol — governed by a foundation or a working group or a consortium of contributors, with no single entity able to change the spec unilaterally — does not have this failure mode. The protocol defines how tasks are described, how workers are matched, how compensation is calculated and delivered. The actual marketplace implementations sit on top of the protocol. Multiple competing marketplaces, any of which can receive tasks from any agent that implements the standard.

This is exactly how the internet's foundational protocols work. HTTP does not care which web server you run. SMTP does not care which email client sends the message. DNS does not care which registrar holds the domain. The protocol is the neutral layer. The implementations compete on quality.

Task-bridge needs to be that neutral layer for human work in AI pipelines. If it is, it can scale to absorb a meaningful fraction of the displaced work that AI creates. If it is captured by any single company, it becomes another platform that extracts value from workers rather than directing value to them.

A direct call — who needs to read this

This is not a think-piece. It is a call to action addressed to specific people building specific things.

The world we are trying to build toward

The article that preceded this one — After Work — ended with Hawking's warning: the outcome depends on how the wealth is distributed, not on how much wealth AI creates. A successful task-bridge protocol is a mechanism for distribution — not of wealth directly, but of work. Of meaningful activity. Of the dignity that comes from being genuinely useful rather than technically supported.

Success looks like this: an AI agent working on a legal contract, a medical record, a financial analysis, a piece of creative work encounters a decision point where it is not confident it can get the answer right. Instead of hallucinating, instead of failing, instead of producing a plausible-sounding wrong answer, it calls the protocol. A human with the relevant skills — a paralegal, a medical coder, a financial analyst, a writer — receives the task, completes it, and is paid fairly. The AI pipeline continues. The human has work.

Multiply that by the millions of tasks per day that AI systems will encounter at their confidence limits. Multiply it by the number of workers who will be displaced from their primary employment by AI automation over the next decade. The match is not perfect — the tasks that AI cannot complete reliably are not always the tasks that displaced workers are immediately qualified for. But it is a start. It is a structural mechanism rather than a political hope. And it can be built now, by the people who are already building the AI systems that create the displacement in the first place.

The honest version

Task-bridge does not solve the job displacement problem. It is worth being precise about this.

The volume of work that AI systems will route to humans through a confidence threshold mechanism is not going to replace the volume of work that AI systems will absorb from human workers over the next decade. The IMF's estimate that 60% of jobs in advanced economies are already exposed to AI disruption is not addressable by routing edge cases to human workers. The scale is different by orders of magnitude.

What task-bridge can do is three things. First, it can make AI pipelines more reliable by replacing hallucinated outputs with genuinely human-completed work at the edges of AI capability. Second, it can create a new category of skilled digital work — AI pipeline oversight, edge-case completion, AI output verification — that represents genuine employment rather than a political concession. Third, it can establish a precedent: that AI systems should be designed with explicit, compensated human handoff mechanisms, not just implicit failure states.

The precedent is arguably the most important of the three. A world where AI pipelines routinely route work to humans and pay them fairly for it is a world where the political economy of AI displacement is different — where AI companies are structurally invested in the welfare of human workers rather than structurally indifferent to them. That is not everything. But it is something real, and it is something that developers building agentic systems can choose to create, right now, by the technical decisions they make in the systems they are already building.

The protocol is at github.com/Jotto1988/task-bridge. The conversation starts there.