Finding
Codebook Agent (arXiv:2609.02264, submitted 2026-09-02) challenges three common assumptions in adaptive multi-agent topology design.
The originating team reports that useful graph searches collapse to about six topology families even when codebook capacity grows from 8 to 64, that graph edge count is negatively correlated with measured token consumption at roughly Pearson r = -0.4, and that profile-node message-passing scorers can become adjacency-invariant when agents share profiles. Their 16-entry codebook emits a topology in 2.4 ms, reaches 84.6 average versus 83.0 for the strongest compared designer, and uses 21.9% to 33.2% fewer LLM tokens. Treat these as originating-team results until reproduced.
RuV implication
Do not optimize Ruflo topology for edge count. Optimize measured task utility, total communication tokens, latency, and failure risk.
Ruflo already has orchestration, agent roles, communication, model routing, and telemetry. The missing experiment is whether real RuV workloads occupy a small reusable topology basis that can be selected without iterative graph search.
Proposed experiment
Capture successful Ruflo communication graphs from representative workloads and normalize them into a topology artifact whose nodes are role identities and whose edges encode allowed communication direction and message class.
Compare:
- current Ruflo topology
- simple deterministic topology families selected by task class
- offline clustered topology codebook with nearest or scored selection
- learned query-to-code selector only if condition 3 exposes a real gap
Start with deterministic clustering before adding a neural selector.
Required measurements
Task success, total model tokens, communication tokens, topology selection latency, end-to-end latency, tool calls, model cost, failure recovery, fanout, retries, and per-agent idle time.
Report graph edge count only as descriptive metadata, never as the optimization objective.
Security
A selected topology controls communication connectivity, not execution authority. It cannot expand RVM capabilities, cross tenant boundaries, bypass delegation scope, or allow peer evidence to become authority. Every topology artifact must identify its allowed principals, message classes, and maximum fanout.
Test malicious or malformed topologies, cycles, disconnected required agents, excessive fanout, topology substitution, stale codebook entries, and role aliasing.
Falsification
Reject the new abstraction if a small hand-written topology set performs within variance, if topology reuse is low, if token savings come from lower success, or if selection overhead erases savings on short tasks.
Acceptance
Advance only if a reusable codebook matches or improves held-out task success while reducing total LLM tokens by at least 20%, or improves held-out success by at least 5 absolute points at matched model cost. Topology selection p95 must stay below 5 ms locally, protected task slices may not regress by more than 2 absolute points, and authority expansion must remain zero.
Coordinate through MetaHarness and Core Memory federation. No autonomous merge.
Finding
Codebook Agent (arXiv:2609.02264, submitted 2026-09-02) challenges three common assumptions in adaptive multi-agent topology design.
The originating team reports that useful graph searches collapse to about six topology families even when codebook capacity grows from 8 to 64, that graph edge count is negatively correlated with measured token consumption at roughly Pearson r = -0.4, and that profile-node message-passing scorers can become adjacency-invariant when agents share profiles. Their 16-entry codebook emits a topology in 2.4 ms, reaches 84.6 average versus 83.0 for the strongest compared designer, and uses 21.9% to 33.2% fewer LLM tokens. Treat these as originating-team results until reproduced.
RuV implication
Do not optimize Ruflo topology for edge count. Optimize measured task utility, total communication tokens, latency, and failure risk.
Ruflo already has orchestration, agent roles, communication, model routing, and telemetry. The missing experiment is whether real RuV workloads occupy a small reusable topology basis that can be selected without iterative graph search.
Proposed experiment
Capture successful Ruflo communication graphs from representative workloads and normalize them into a topology artifact whose nodes are role identities and whose edges encode allowed communication direction and message class.
Compare:
Start with deterministic clustering before adding a neural selector.
Required measurements
Task success, total model tokens, communication tokens, topology selection latency, end-to-end latency, tool calls, model cost, failure recovery, fanout, retries, and per-agent idle time.
Report graph edge count only as descriptive metadata, never as the optimization objective.
Security
A selected topology controls communication connectivity, not execution authority. It cannot expand RVM capabilities, cross tenant boundaries, bypass delegation scope, or allow peer evidence to become authority. Every topology artifact must identify its allowed principals, message classes, and maximum fanout.
Test malicious or malformed topologies, cycles, disconnected required agents, excessive fanout, topology substitution, stale codebook entries, and role aliasing.
Falsification
Reject the new abstraction if a small hand-written topology set performs within variance, if topology reuse is low, if token savings come from lower success, or if selection overhead erases savings on short tasks.
Acceptance
Advance only if a reusable codebook matches or improves held-out task success while reducing total LLM tokens by at least 20%, or improves held-out success by at least 5 absolute points at matched model cost. Topology selection p95 must stay below 5 ms locally, protected task slices may not regress by more than 2 absolute points, and authority expansion must remain zero.
Coordinate through MetaHarness and Core Memory federation. No autonomous merge.