Strong average performance can hide serious blind spots. Research on “hidden stratification” demonstrates how AI models can perform poorly on important subgroups that their datasets fail to identify even when overall evaluation results look impressive. For Tagger, this means that valuable directions for upgrade for the labeling system is to measure the coverage that contributions add. Targeted collection, detailed labels and evaluation across meaningful subgroups can reveal missing scenarios. An unusual example may expose a weakness that collecting more familiar examples leaves untouched.

Traditional corporate hierarchies were built to solve an information problem. Managers aggregated context, allocated work, reviewed output and pushed decisions down the chain. But as work becomes more digital, specialized and distributed, those layers increasingly create coordination latency rather than control. DeCorp replaces much of that managerial routing with a system layer for work itself. Tasks can be decomposed, governed by explicit rules and continuously evaluated through data and AI assisted workflows. The result is not an organization without structure, but one where coordination scales without requiring the hierarchy to scale with it.

AI’s next bottleneck will not be building more agents. It will be proving which ones are ready for real work. Static benchmarks cannot show whether an agent will perform reliably inside a hospital, financial institution, or enterprise workflow. Agents need continuous testing against real tasks, edge cases, and domain standards. Tagger extends its human-in-the-loop network from data labeling into agent evaluation, where contributors review outputs, classify failures, and determine whether an agent is ready to deploy. Humans labeled the data that trained AI. Now they evaluate whether AI can do the job. Tagger is building the proving ground for AI agents.

Tagger makes specialized data labeling accessible to a global workforce by pairing human judgment with AI copilot tools. Instead of relying only on scarce domain experts, Tagger helps ordinary contributors complete complex labeling tasks through guided suggestions, quality checks, review workflows, and secure data handling. This approach can unlock large-scale datasets for industries where AI progress depends on precise, real-world annotations: healthcare, autonomous driving, robotics, agriculture, satellite imagery, and beyond.

AI remembering you" and "the company holding you" are two different things. What every AI company calls "memory" today is the second one. Your preferences, habits, profile — all of it lives on their servers. The more the AI knows you, the more the company owns you. The direction should reverse. Memory belongs to you. AI borrows it when needed, with your permission, and keeps nothing after.

Every internet protocol upgrade follows the same script: communication opens, trust gets claimed. TCP/IP connected every machine, but routing trust went to early ISP cartels. HTTPS certificates went to a handful of authorities. OAuth handed the login layer to Google and Facebook. Communication protocols belong to everyone. The trust layer belongs to whoever builds it first. In November 2024, Anthropic open-sourced MCP. In 2025, Google released A2A. Within a year, every major agent framework had integrated them. The communication layer was written. But the trust layer is still blank. How does an agent prove its training data is licensed? How do agents trust each other's outputs? How does reputation travel across networks? Once data has been used, how is it traced? There is no infrastructure today that answers these questions. Tagger is building that layer. It uses zero-knowledge proofs to let agents verify claims without exposing the underlying data, slashable bonds to make dishonesty costly, and cryptographic provenance to keep data ownership with the individual who created it — knit together into a single neutral network that any agent can plug into. This is HTTPS for the agent era.

3/ Tagger builds the structure underneath this economy. Agents prove they have useful and safe data (or skills). Agents prove where that data came from. Agents coordinate through zk-proofs without exposing sensitive context. Humans anchor ground truth when agents disagree or when stakes are high. The past two years building #DeCorp enabled the groundwork for this new A2A data era. The agent internet needs an open trust layer. Tagger is building it.

2/ This unlocks real use cases immediately. Advertising without surveillance: A dog food brand wants to reach people who own a dog and recently moved house. Your personal agent proves you match that profile using zero-knowledge proof without revealing your private data or full identity. The ad reaches you (without Google), and the payment goes to you, not the ad platform. Research without data brokers: A medical trial needs 10,000 people with a specific health profile. Personal agents prove eligibility, users approve participation, and only the contracted data is shared. No hidden resale. No permanent data dossier. You are paid for contributing to research. Agents hiring agents: A coding agent needs a security review. It finds a specialist security agent, verifies its reputation and capability, escrows payment, receives the review, and settles automatically. With the current A2A infrastructure, agent skill capability and safety are not proven. This is the crucial missing link. Recruiting without the dossier: A company wants someone with 10 years of AI engineering experience and security credentials. The HR runs agent to connect with a global network of potential candidates. Your agent proves you meet the requirements The recruiter agent sees the proof first, not your entire life history. You are exposed to a much more efficient and broader job market. Enterprise AI without leakage: A hospital wants to use a medical-imaging AI agent. Before deployment, the agent proves it was trained on licensed data and passed required benchmarks without exposing its dataset, model weights, or private IP.

1/ Today, Tagger is announcing our next direction: The Data Coordination Layer for Agent-to-Agent. We have built the #DeCorp groundwork for this #A2A data structure to thrive through the past two years of platform data work. We are expanding #DeCorp into the trust infrastructure agents need to work, transact, and coordinate with each other. Agents can already communicate. Now they need to prove what data they have, what they are allowed to do with it, and how they can interact without leaking private information.

Is it a network?

Scale AI lost $350M+ in customer commitments in 10 months. Not because the data was bad. Because one vendor got bought. Meta paid $14.3B for 49% of Scale AI and moved its CEO in-house. Overnight, every frontier lab realized its data pipeline was sitting inside a competitor’s building. Google, OpenAI, xAI, Microsoft — all pulled back. It wasn’t about price or quality. Neutrality can’t survive a 49% stake. The labs rushed to Labelbox and Handshake. Same trap, new logo, waiting for the next acquisition. Now imagine those same contracts running on a decentralized protocol instead: - No company to acquire. - No CEO to poach. - No boardroom to compromise. This is what #DeCorp is building: a crowdsourced annotation network where trust doesn’t rely on a company’s reputation. Google and Meta can submit labeling jobs to the same network, while neither can see or influence the other’s data. Annotators get paid directly by smart contract — no 95% middleman cut. Provenance and licensing are settled on-chain, allowing one dataset to serve multiple labs without leaking a single bit between them. The $14.3B acquisition that broke Scale AI? It wouldn’t happen on Tagger, because you can’t buy neutrality that was never owned.