Make the structure of an agent system something we can study.
Agent Graph is a research direction at Nuclx. It represents agents, tools, memory, and evaluators as a computational graph, then asks whether its topology, feedback mechanisms, and evaluation methods can be improved systematically.
One graph describes how a collaboration unfolds.
We model an agent-system run as a directed computational graph. Nodes hold capabilities or state. Edges direct information, work, and feedback. A complete run produces an inspectable trace.
- Agents
- Model-driven units for reasoning, planning, or execution, each with explicit inputs, outputs, and responsibilities.
- Tools
- Callable capabilities that connect code, data, and external systems, with permissions, failure states, and costs made explicit.
- Memory
- Task state, prior traces, and reusable knowledge. The policies for reading and writing memory also affect system behavior.
- Evaluators
- Components that test intermediate work and final outcomes against predefined criteria, then return feedback to the graph.
Not only the prompt. The structure, too.
-
Topology
Which nodes participate, how they connect, and when work branches or converges. Every change should be recorded, compared, and reversible.
-
Feedback
How evaluation signals return to a run—and whether they revise the current trace, update a policy, or propose a different graph.
-
Evaluation
Task quality, cost, latency, reliability, and generalization considered together, so a single score does not hide the real trade-offs.
Validation method
Make the question reproducible before discussing evolution.
An optimizable graph is not an explanation of intelligence.
- An evaluator is only a proxy for the intended objective. A higher score may reflect exploitation of the evaluation rule, not a genuine capability gain.
- Model versions, sampling parameters, tool availability, and changing data can confound comparisons; experiments must control and disclose them.
- More complex topology adds cost, latency, and failure paths. Graph size is not a research objective in itself.
- Systems with permissions, external actions, or high-stakes decisions require human oversight and explicit stopping conditions.
Starting with representation and measurement.
Nuclx is currently focused on the graph representation, execution traces, and evaluation protocol. We will only state findings after experiments are reproducible, baselines are clear, and held-out validation has been completed.