The DAG of AI
Every working AI system is a graph of contracts.
Transcoders are the atoms. Forges are the molecules. Contracts govern both. The result is a directed acyclic graph — and every effective AI system converges on this shape.
The shape
Strip away the branding from any AI system that actually works — grocery ordering, security auditing, content publishing, enterprise procurement — and you find the same structure. Nodes that do work. Edges that carry artifacts. Contracts that define what "done" means at each node.
This isn't a design choice. It's a convergence. Systems that don't converge on this shape break when the model changes, when the team scales, or when the requirements shift. Systems that do converge survive all three.
The shape is a directed acyclic graph. A DAG.
Atoms and molecules
A transcoder converts content from one form to another. A document becomes a podcast. A YAML contract becomes a PDF. A briefing becomes a slide deck. Each transcoder has a contract that defines what the output must be — not how to produce it. The contract is durable. The model that fulfills it is replaceable.
A forge chains transcoders into a pipeline. A podcast forge runs a scriptwriter, then a voice renderer, then an audio producer, then mastering. Each node is a transcoder. Each edge is an artifact. The forge is the graph.
Eidos Human
Contracts govern both levels. The forge has a contract (what the finished podcast sounds like). Each transcoder has a contract (what a valid script looks like, what a clean audio render sounds like). The contracts compose. The models are interchangeable at every node.
Two kinds of output
DAG nodes produce artifacts. But not all artifacts are the same.
- Consumed by humans
- Podcasts, PDFs, emails, grocery orders
- Pipeline ends
- Value is in the output
- Leaves of the graph
- Consumed by agents
- Schemas, DAGs, contracts, state files, memory
- Pipeline continues
- Value is in making the system smarter
- Edges of the graph
A podcast episode is a terminal artifact. The listener consumes it. A project DAG is a persistent artifact — the agent that built it left behind a map that makes every future agent session better. The forge didn't just do work. It built infrastructure.
The protocol layer
The internet is a DAG of protocols fulfilled by interchangeable machines. TCP/IP, HTTP, DNS — contracts that any server can fulfill. The machines are replaceable. The protocols are permanent. The web won because it got the contracts right, not because it got the servers right.
Eidos Human
Nobody remembers which server served them a web page. Everyone depends on DNS. Nobody will remember which model fulfilled a contract. Everyone will depend on the contract that defined what a valid output looks like.
Eidos is building the contract protocol layer for the agentic internet.
The third leg
Minimum Viable Data says data is the primitive, not product. Contracts, Not Skills says define what, not how. The DAG says these compose into graphs that any agent can traverse.
Together they form a complete architecture for AI systems that get better as models improve, rather than breaking.
- MVD — the data layer is permanent
- Contracts — the specification layer is durable
- DAGs — the composition layer is emergent
The agent reads the contracts, discovers the graph, and traverses it. No human draws the workflow. The shape emerges from the constraints.
What this replaces
Hand-configured workflows. n8n, Zapier, Make, GitHub Actions for AI orchestration, LangChain prompt chains. Every one of these requires a human to draw the graph. When requirements change, someone redraws it.
In a contract-driven system, the agent discovers the graph by reading the contracts. When the contracts change — new compliance requirements, new cost thresholds, new output formats — the agent adapts the pipeline. No one rebuilds the workflow. The graph reconfigures itself because the constraints reconfigured.
This is autonomous workflow. Not automated workflow. The difference is who draws the graph.