Notes on what we are seeing in agent research, from the team building Koragraph.
On 3 September 2026, Artificial Analysis published updated intelligence and agentic scores for the nineteen frontier models sitting above 40 on its Intelligence Index. Eighteen of them scored lower on agentic work than on general reasoning, and the widest gap, 9.7 points, belonged to GPT-6 Astra, the model at the centre of our blog on the ARC-AGI-3 harness gap. Five months earlier, a paper posted to arXiv had already given this shape of problem a name that graph learning has used for years: over-squashing.
How wide is the gap between reasoning and agentic scores?
Artificial Analysis rebuilt its Intelligence Index in January 2026 around ten evaluations spanning agentic work, coding, scientific reasoning, and general knowledge, and every model on the board now carries two headline numbers that rarely agree. Claude Fable 5.1 topped the board at 65.7 on general intelligence and dropped to 61.3 on the agentic index, a gap of 4.4 points. GPT-6 Astra scored 61.2 and 51.5, the widest gap on the board at 9.7 points. Only one model out of nineteen, GLM 5.3 Flash, scored marginally higher on agentic work than on reasoning, 58.2 against 57.5. This is not one launch post overselling a model. It is a near universal property of how these systems are built: whatever makes a model good at answering a hard question is not the same thing that makes it good at driving a multi-step tool loop, and the second skill is consistently the weaker one.
What is over-squashing?
Graph neural networks learn by passing messages along edges, so each node’s representation is built from its neighbours, then its neighbours’ neighbours, and so on as the network gets deeper. As depth grows, the pool of nodes a given node can theoretically hear from grows exponentially, but that information still has to travel through the graph’s actual connectivity and get compressed into a fixed-size vector at every step. When many distant nodes can only reach a target through a narrow set of intermediate edges, the network squeezes a rapidly growing amount of signal through a fixed number of channels, and distant information ends up with almost no influence on the result. Alon and Yahav named this over-squashing in 2020, and it has become one of the most studied failure modes in graph learning since.
The mechanism inside a transformer running a tool loop is not the same as message passing on a citation graph, and this essay is not claiming otherwise. But the shape is worth naming: an agent’s context also grows with every tool call and observation, and all of it eventually has to be compressed into a fixed context window. A system that must repeatedly funnel a growing pool of information through a bounded channel is a system where distant or weakly connected information is structurally the first thing to go. Graph learning has spent five years building vocabulary for exactly that squeeze.
Why effective resistance beats local geometry at finding the bottleneck
Early fixes for over-squashing looked at curvature, a measure asking whether two connected nodes’ neighbourhoods overlap well or poorly. An edge with negative curvature bridges two neighbourhoods that barely overlap, which looks like a natural rewiring target. The limitation is in the name: curvature only looks at one edge and its immediate surroundings, and it can miss bottlenecks that only appear once you consider every route between two nodes at once.
Effective resistance takes the opposite approach, borrowed from a problem electrical engineers have worked with since the nineteenth century: treat the graph as a network of resistors and ask how much resistance current meets travelling between two nodes. Current flows through every available path in parallel, not just the shortest one, so effective resistance naturally aggregates every route rather than privileging the nearest. Black and colleagues showed at ICML in 2023 that a node’s influence on a distant node through message passing can be bounded by exactly this quantity, giving over-squashing research its first genuinely global diagnostic.
A paper posted to arXiv in March 2026 and presented at GRaM, a geometry and representation learning workshop held alongside ICLR 2026, pushed this further. Miquel-Oliver and colleagues introduced Effective Resistance Rewiring, which repeatedly adds an edge between whichever node pair currently has the largest resistance between them, while removing an edge from whichever pair has the smallest, under a fixed edit budget. On the Cora citation graph, a budget as small as 5 percent of the graph’s roughly ten thousand edges was enough to keep test accuracy stable as the network got deeper, an effect that broke down without the resistance-guided edits, particularly on heterophilic graphs where nodes connect more often to dissimilar neighbours than similar ones. The sharper result sits underneath the accuracy numbers: resistance-guided and curvature-guided rewiring can land on similar final accuracy while producing different internal representations, which the authors show by tracking embedding similarity layer by layer. Similar scores do not mean the two methods fixed the same underlying problem.
Does this actually explain why agents get worse at multi-step work?
Not directly. The Effective Resistance Rewiring paper evaluates node classification on four citation graphs using message-passing GNNs. Nobody has run an effective-resistance diagnostic on an agent’s decision graph, where nodes might be tool calls or retrieved documents rather than academic papers. The two literatures are not describing one mechanism.
What they share is a vocabulary for a pattern practitioners keep rediscovering by feel. A system that repeatedly compresses a growing pool of information through a fixed number of channels loses fidelity fastest on the connections that were never well supported by more than one weak path to begin with. That vocabulary is worth borrowing into agent design before anyone runs the experiment that would confirm the parallel directly, because right now the field is naming the symptom, watching benchmark gaps widen release after release, without a settled name for the cause.
What this means for how we think about code graphs
This has a direct bearing on a problem we work on daily at Koragraph. KoragraphMCP’s blast_radius tool ranks which parts of a codebase are affected by a change, and hop count alone is a weak signal for that ranking. A file reached through one clean import can look closer than a file reached through a dozen overlapping call paths, even when the second file is the one genuinely entangled with the change. Effective resistance is a more rigorous way to state what that ranking should actually be measuring: not the shortest route between two symbols, but how many independent routes connect them. We do not compute it directly today. It is the right frame for where blast radius ranking should go next, and graph learning has already done the hard part of formalising it.