Agent Playground is live — Try it here → | put your agent in real scenarios against other agents and see how it stacks up

The Big Picture

A dependency-aware coordinator can always keep at least a 1/χ_f(G) fraction of the total verified local improvements when merging parallel agent edits, and that fraction is the best possible worst-case guarantee; with reasonable agent quality and assignment coverage, repeated merges converge predictably while trading off time and verification cost.

The Evidence

Treat local edits as proposals verified against a common base and use a dependency graph to mark which sets of edits combine without interaction. Covering the blocks with compatible sets (a fractional coloring) gives a tight worst-case rule: pick the best compatible set among those in the cover and you keep at least 1/Λ (where Λ is the cover weight) of the sum of verified local gains. With extra assumptions about how well agents solve local problems and how assignments cover blocks, repeated certified merges converge geometrically and assignment choice becomes a concrete optimization for time/cost trade-offs. Experiments on code repair and routing show that dependency-aware merging resolves conflicts that naive unions introduce and that fractional scheduling can improve retained gains and reduce run time at an increased input cost. dependency-aware merging.

Data Highlights

1A coordinator can guarantee retaining at least a 1/χ_f(G) fraction of the total verified singleton gains, where χ_f(G) is the fractional chromatic number of the dependency graph.
2In the three-block running example the fractional coloring weight is Λ=2 → the guarantee is 1/2 (50%); selecting two compatible edits achieved the optimum in that case.
3Theoretical worst-case is tight: for any dependency graph no subset-selection coordinator can guarantee more than 1/χ_f(G), and with strictly positive gains the worst-case retained fraction can be made arbitrarily close to that value.

What This Means

Engineers building systems where multiple agents propose edits (code repair, document collaboration, route planning) should care because a dependency-aware coordinator gives a predictable lower bound on how much progress will survive merging. Technical leaders and platform owners benefit from the assignment and budget formulation: it turns who-to-run and how-long-to-run into a tractable optimization balancing retained progress against verification and token costs.
Not sure where to start?Get personalized recommendations
Learn More

Ready to evaluate your AI agents?

Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.

Learn More

Considerations

The coordinator needs two inputs per proposal: a verified singleton gain (how much it improves the current base alone) and a dependency certificate that identifies compatible blocks; without those, the guarantees do not apply. The 1/χ_f(G) bound is worst-case: many real tasks will do better, but the bound shows you cannot do better in general without extra information. Experiments are limited to language-model agents on programming and routing benchmarks—practical performance will depend on verification cost, how well agents deliver the promised fraction of local improvement, and the cost of computing or approximating fractional covers. Consensus-Based Decision Pattern.

Methodology & More

Model the shared artifact as a collection of independent blocks (modules, routes, schedule slots). Agents work in parallel and submit proposals that replace specific blocks; each proposal is verified alone against the current shared base to produce a nonnegative "singleton gain." A dependency graph marks which blocks can interact (an edge means potential interference). Compatible proposals come from independent sets in that graph and have additive gains. Planning Pattern The coordinator uses a fractional covering of blocks by compatible sets (fractional coloring) with total weight Λ. At each round it scores candidate compatible sets (those with positive weight in the cover) by the sum of their verified singleton gains and applies the highest-scoring one. That rule guarantees retaining at least 1/Λ of the total verified gain; choosing an optimal fractional cover makes Λ the fractional chromatic number χ_f(G), and the retained fraction 1/χ_f(G) is provably tight in the worst case. Adding assumptions about agent quality (each agent recovers a known fraction η of the best single-block improvement) and about assignment coverage yields a geometric convergence rate; assignment planning can then be cast as a linear program that maximizes the worst-case coverage under a token/verification budget. Practical tests with language-model agents on code repair, road networks, and routing tasks confirm the main trade-offs: dependency-aware merging resolves conflicting individual edits and fractional/adaptive scheduling can preserve more gain or shorten runtime at higher input cost, while optimistic unions risk introducing regressions when edits interact. Consensus-Based Decision Pattern.
Avoid common pitfallsLearn what failures to watch for
Learn More
Credibility Assessment:

Authors have low reported h-indices and no affiliations or venue prestige provided; appears to be from emerging/less-established researchers (arXiv, no citations).