Information about AI agents is fragmented. Capabilities live on product pages. Model and protocol support is buried in documentation. Benchmark results appear on separate leaderboards. Risks are often copied from one directory to another. Price is frequently missing altogether.

ATG — the Agent Taxonomy Graph — turns these disconnected claims into a structured, public knowledge layer. It maps 133 agent systems through typed relationships to capabilities, production controls, protocols, models, runtimes, risks, benchmarks, maintainers, and sources.The goal is to make every claim inspectable.

agent → capability → model → protocol → benchmark → evidence

In ATG, every relationship is also a claim. Claims can carry a source, evidence, a verification timestamp, and a confidence level:

Agent systems, documentation, model support, and benchmark results change. ATG is designed as an update loop: discover a claim, check its evidence, assign confidence, and re-audit it when the underlying source changes. The graph records not only what we know, but how well we know it and when it was last checked.

Geo provides canonical entities, reusable typed relationships, public provenance, and a shared knowledge layer that other people can inspect and extend. Instead of creating another isolated copy of a model, agent, benchmark, or source for every publication, new work can connect to the same public entities. ATG becomes public infrastructure that other publications, applications, and researchers can build on.

The current dataset aggregates 131 evaluation runs: 46 from Princeton HAL; 73 usable Terminal-Bench 2.0 evaluation records; 12 curated observations from public sources.

The strongest example appears when we hold the benchmark task and external verifier fixed. The task pytorch-model-cli has 47 evidence-linked successful passes. The cheapest recorded pass cost: $0.0013227 — MiniMax M2.7 The most expensive recorded pass cost: $1.4668124 — Claude Opus 4.6 Both passed the same external verifier. Same task. Same verified outcome. 1,109× price difference. The ratio itself can be calculated in a table. The value of the graph is that every part of the result remains connected:

task → successful pass → model → benchmark → source → recorded cost

This is where agents become economically important. A model provides capability, but the agent or harness decides how that capability is used: which model is sufficient for the task, when to escalate, and when to stop after verification. A good agent should not send every task to the most expensive model. It should find the lowest-cost configuration that can complete the task reliably.

ATG is not another static list of AI agents. It is a public, verifiable layer for understanding how the agent ecosystem fits together — and eventually, how agents can choose the right model for the right task at the right cost. Inspect an edge. Follow its evidence. Challenge a confidence level. Build another view. The graph is open.

The spread is the story. Average cost per evaluated configuration ranges from $0.03 to over $1,600 — five orders of magnitude for systems sitting side by side on the same leaderboards. A $0.06 configuration already clears a 60% average score; the last twenty-five points cost 10–30× more. At the level of a single task the market is wildly unpriced: pytorch-model-cli was solved by 47 different agent×model configurations for anywhere between $0.001 and $1.47 — the same outcome at a 1,109× price difference. Even a fixed agent+model pair varies several-fold run to run: every trial takes its own trajectory. Explore the live map: Task Prices v2 on the ATG dashboard.