Attractor is our implementation of a non-interactive coding agent.
It composes models, prompts, and tools into a graph-structured pipeline. It is designed to operate end-to-end once the work is fully specified.
Graph Structure
"Attractor is structured as a graph of nodes, forming a generative SDLC.
Each node corresponds to a phase of work and is governed by a core prompt (e.g. "implement the functionality", "identify the bottleneck").
Example Nodes
I
Implement
"Implement the functionality"
I
Identify
"Identify the bottleneck"
O
Optimize
"Optimize for performance"
V
Validate
"Verify behavioral correctness"
Natural Language Edges
Edges between nodes are expressed in natural language and evaluated by the LLM.
Example Edges
Proceed Edge
"Proceed once a bottleneck is identified"
Judged Branch
"Take this edge if the copywriting standards have been met"
"Take this edge if we need to request an exception to the standards"
Execution Model
Execution consists of traversing this graph until convergence or termination conditions are met.
S
→
I
→
O
→
V
→
✓
Start → Implement → Optimize → Validate → Complete
Key Properties
- Deterministic given the same inputs
- Observable at every node transition
- Resumable from any checkpoint
- Composable with other graphs