Andrej Karpathy just dropped 12-page PDF on "Graph Engineering" for multi-agentic systems
the shift: Karpathy's loop runs 700 experiments and forgets all of them. A graph remembers forever
here's the full system:
step 1 → build one loop: generate, critique, revise. 630 lines, 700 experiments in 48 hours
step 2 → go parallel: agents in separate worktrees, same repo, different branches, no conflicts
step 3 → add a knowledge graph: extract entities, resolve aliases, assemble typed edges, query through subgraphs
step 4 → ground your evaluator: it checks claims against graph edges, not vibes
step 5 → plug the graph as shared memory. workers write to it. evaluators fact-check against it. Loops persist overnight
step 6 → the agent forgets. the graph does not. stop rebuilding context from scratch every session
Karpathy ran 1 agent in 1 direction. Anthropic's graph runs 1,000 with shared memory - same model, it's the architecture
this 11-page PDF changed how I'm building multi-agent systems today
read it now - then explore the full graph engineering article below ↓
the shift: Karpathy's loop runs 700 experiments and forgets all of them. A graph remembers forever
here's the full system:
step 1 → build one loop: generate, critique, revise. 630 lines, 700 experiments in 48 hours
step 2 → go parallel: agents in separate worktrees, same repo, different branches, no conflicts
step 3 → add a knowledge graph: extract entities, resolve aliases, assemble typed edges, query through subgraphs
step 4 → ground your evaluator: it checks claims against graph edges, not vibes
step 5 → plug the graph as shared memory. workers write to it. evaluators fact-check against it. Loops persist overnight
step 6 → the agent forgets. the graph does not. stop rebuilding context from scratch every session
Karpathy ran 1 agent in 1 direction. Anthropic's graph runs 1,000 with shared memory - same model, it's the architecture
this 11-page PDF changed how I'm building multi-agent systems today
read it now - then explore the full graph engineering article below ↓
Great idea 🔥
Here’s a more complete and robust prompt for Claude (works especially well with Projects or when you upload the codebase):
Analyze my entire codebase in depth.
Generate TWO complete, ready-to-use deliverables:
1. A single self-contained HTML file (Tailwind CDN + dark theme) that includes:
- Interactive architecture diagram (nodes + edges)
- Flows panel on the right
- When selecting a flow, highlight the full path in the diagram and show detailed steps below
- Tooltips with descriptions for each component
- Clean, professional, and responsive design
2. A structured JSON with this exact shape:
{
"nodes": [...],
"edges": [...],
"flows": [
{
"id": "...",
"name": "...",
"description": "...",
"steps": [...]
}
]
}
The HTML is for humans. The JSON is for the next AI agent so it can fully understand the architecture and work on new features without losing context.
Deliver both files complete, with nothing omitted.
Here’s a more complete and robust prompt for Claude (works especially well with Projects or when you upload the codebase):
Analyze my entire codebase in depth.
Generate TWO complete, ready-to-use deliverables:
1. A single self-contained HTML file (Tailwind CDN + dark theme) that includes:
- Interactive architecture diagram (nodes + edges)
- Flows panel on the right
- When selecting a flow, highlight the full path in the diagram and show detailed steps below
- Tooltips with descriptions for each component
- Clean, professional, and responsive design
2. A structured JSON with this exact shape:
{
"nodes": [...],
"edges": [...],
"flows": [
{
"id": "...",
"name": "...",
"description": "...",
"steps": [...]
}
]
}
The HTML is for humans. The JSON is for the next AI agent so it can fully understand the architecture and work on new features without losing context.
Deliver both files complete, with nothing omitted.
Analyze my codebase. Generate a self-contained HTML (Tailwind CDN, dark theme) with an interactive architecture diagram + interactive Flows section (select and highlight path + steps).
And a JSON {nodes, edges, flows: [{steps}]} for AI agents.
Deliver both complete.
And a JSON {nodes, edges, flows: [{steps}]} for AI agents.
Deliver both complete.
How to become an AI-native company:
1. Establish a unified Model Context Protocol (MCP) or API gateway to enable team connectivity with internal systems.
2. Develop a centralized corporate knowledge repository:
- Integrate static context data, including organizational identity, operational guidelines & product documentation.
- Integrate dynamic context data, including meeting minutes, email correspondence, Slack communications, and active project details.
3. Implement a corporate orchestration framework to instruct AI on interacting with the knowledge repository and internal systems.
4. Integrate the workforce into the orchestration framework to establish a self-improving feedback loop that learns from employee activities.
5. Develop a model-routing layer to evaluate and distribute tasks to the appropriate model at the optimal time, thereby mitigating vendor risk.
6. Construct autonomous agents atop the corporate orchestration framework.
1. Establish a unified Model Context Protocol (MCP) or API gateway to enable team connectivity with internal systems.
2. Develop a centralized corporate knowledge repository:
- Integrate static context data, including organizational identity, operational guidelines & product documentation.
- Integrate dynamic context data, including meeting minutes, email correspondence, Slack communications, and active project details.
3. Implement a corporate orchestration framework to instruct AI on interacting with the knowledge repository and internal systems.
4. Integrate the workforce into the orchestration framework to establish a self-improving feedback loop that learns from employee activities.
5. Develop a model-routing layer to evaluate and distribute tasks to the appropriate model at the optimal time, thereby mitigating vendor risk.
6. Construct autonomous agents atop the corporate orchestration framework.
