Draft specification 0.1.0

DAG-TOML proves what an agent intended, depended on, and made ready for review.

A TOML format for declarative policy, workflow, traceability, evidence, closure roots, and review gates in agent-driven software development.

Read the spec GitHub llms.txt

Diagram showing intent, dependency graph, evidence, validators, and review decision flow. — Evidence layer, not an execution runtime.

Positioning

The evidence layer for agentic software work

Declares intent

Agents describe planned work, ordered dependencies, traceability, evidence expectations, and review state before the result is treated as ready.

Preserves provenance

Closure-root rules tie claims to upstream source hashes so downstream artifacts can inherit brittleness instead of hiding it.

Stays runtime-neutral

DAG-TOML does not schedule tools, grant permissions, move messages, or sandbox code. Control planes and runtimes keep owning those decisions.

Use it when

Diff review alone is no longer enough

DAG-TOML is aimed at software generated or modified by agents, especially where a reviewer needs to evaluate the claimed shape of the work before reading every byte.

Agent-generated implementation plans and dependency graphs
Traceability from requirement to code, test, output, and evidence
Human review gates for self-modifying or high-impact changes
Disclosure, redaction, and selective evidence release
Cost accounting as a first-class artifact

Start points

Human and machine-readable entry points

Normative specRoot metadata, template kinds, closure roots, IJB, and extension rules. ProfilesAgent Assurance, disclosure, cost, and deployment tiers. ValidatorsRust and Go primary validators, Python reference validators, and local checks. Protocol comparisonHow DAG-TOML relates to MCP, A2A, ACP, ANP, Agent Manifest, and JSON Agents. Agent readinessDiscovery files, Markdown mirrors, structured metadata, and retrieval design. Markdown sitemapCompact navigation for coding agents and retrieval systems.