← graph·atlas · published on Geo Protocol

Agent Taxonomy Graph

A field guide to the AI agent ecosystem — built like a specimen catalog, verified like a scientific record. Every capability claim carries its evidence; every risk is derived, not assumed.

133agent systems
2,600+published relations
4,392evidence rows
19relation types

01 · why

Choosing an agent is still guesswork

Vendor pages overclaim. Leaderboard numbers arrive without harness or date. Directory sites copy-paste the same risk warnings onto every tool. When we audited our own early data, 82 of 133 agents shared one of just three identical risk lists — decoration, not analysis.

ATG's answer is structural: separate what is claimed from what is verified, and never let a completeness score masquerade as a quality score.

02 · what's inside

Every agent is a specimen, fully classified

Agents link to 45 core capabilities, 25 production controls, 14 protocols, 38 models, 24 runtimes, 11 risk types, and 19 benchmarks — through typed, weighted, evidenced edges.

Hermes Agent nousresearch-hermes-agent
ClusterAutonomous generalist
tool_useverified
approval_gatesverified
audit_logneeds review
RuntimeDocker · SSH · Modal · Daytona
LicenseMIT

03 · methodology

Evidence first, confidence second

A claim enters the graph only through the pipeline. Official docs are fetched and content-hashed, snippets extracted with character offsets, decisions logged with rationale — and every change lands in an append-only audit log.

Source documents57 roots · hashed · dated
Evidence snippetschar-offset excerpts
Edge decisions151 reviewed
Audit log378 changes tracked
271
inferred

Rule-seeded from metadata. Visible, but styled as unproven.

140
reasoned

Taxonomy analysis, no direct excerpt yet.

55
verified

Backed by an official-source excerpt with URL and date.

03b · loop engineering

Confidence is a cycle, not a stamp

A static label degrades silently. ATG runs a verification loop: claims enter as inferred, earn evidence through a critic pass, and are re-checked on a schedule — so a verified edge means "verified as of the last audit," not "verified once and forgotten."

Claimdiscovered from source
Fact gatecritic-only pass
Confidenceverified · reasoned · inferred
Re-auditdrift detector, scheduled

The three gates — fact-check on entry, evidence-quality on assignment, calibration on schedule — are what separate a loop with gates from an agent agreeing with itself at speed. The 127 pending evidence decisions aren't a backlog; they're the active queue of this cycle.

04 · risk model

Risks are derived, never templated

A rules engine computes each agent's risk surface from its documented behaviors, then subtracts evidenced controls. An agent that ships approval gates genuinely scores safer than one that doesn't — 952 risk edges, all explainable.

shell_execution + multi_step_autonomy
approval_gates sandboxing audit_log
= active risk profile

05 · benchmarks

Numbers travel with their provenance

No score enters ATG without model, harness, source, and date. Vendor-reported and independently-standardized runs are labeled differently — because they differ by 10–30 points in the wild.

Codex CLI83.4Terminal-Bench 2.1 · GPT-5.5 · tbench.ai
Claude Code78.9Terminal-Bench 2.1 · Opus 4.8 · tbench.ai
Gemini CLI70.7Terminal-Bench 2.1 · Gemini 3.1 Pro · vendor
Mini SWE-agent70.6SWE-bench Verified · Sonnet 4.5 · academic

Frameworks (LangChain, CrewAI, …) carry no score by design — a library is not an end-to-end product.

06 · where it lives

Published on Geo Protocol

ATG ships as a public knowledge graph on Geo — 5,896 graph operations, 337 comparison blocks, 280 support entities. Anyone can browse it, cite it, or challenge an edge through community governance.

Confidence levels map to visual styles in the published graph: verified edges render solid, reasoned dashed, inferred faint. What you see is what we can prove.

07 · next

The catalog keeps growing

Next milestones: close the verification loop on 127 pending evidence decisions through scheduled fact-gate and drift-detector passes, expand the verified core beyond 30 agents, self-run Terminal-Bench evaluations via Harbor, and independent scoring axes — completeness, task-fit, benchmark, risk — that never collapse into a single misleading number.

graphatlas.tech · @maximl · Agent Taxonomy Graph v15 · 2026 · explore the dashboard (v1) →