The Big Picture
Putting each tool in its own process and keeping a single append-only session log on a shared bus lets agent sessions resume after crashes without repeating effects and with almost no runtime penalty.
ON THIS PAGE
The Evidence
Moving plugin assembly and composition out of a single host process and onto a lightweight bus preserves the ability to undo and recover actions across process boundaries. A small router, per-tool processes, and a shared append-only transcript let any process die and be replaced by rebuilding the session from the transcript (cold switching) without redoing already-settled work. Experiments show recovery across many crash points, mixed-language nodes, and heavy concurrency with no lost or duplicated messages and very low per-hop cost compared to model latency. The design also narrows the blast radius of faults: a killed process only affects its own node rather than every co-resident session. cold switching
Data Highlights
13,500 calls from up to 200 concurrent callers were routed with no message loss, duplication, or misattribution in stress tests.
2Twelve end-to-end sessions survived six process kills and resumed through cold switching with no repeated actions; 80 additional sessions resumed after kills at four different tool-call points with no duplicate effects.
3Median bus hop cost was 0.215 ms — about 1/823 of the model's 177 ms first-token latency — making routing overhead effectively invisible at model time scales.
What This Means
Platform engineers building agent systems and tool ecosystems can use this pattern to reduce downtime and avoid whole-system restarts when one tool fails or is upgraded. Technical leaders and reliability teams benefit by shrinking the failure domain: individual tool crashes stop at that process instead of taking down every session in the host. Researchers and evaluators interested in multi-agent trust and agent-to-agent evaluation get a practical architecture for preserving correct action histories across failures. agent-to-agent evaluation
Not sure where to start?Get personalized recommendations
Key Figures
Fig 1: Figure 1. The Logos construction, peer processes on a bus, the router holding only its routing table, harnesses and tools as nodes, and the append-only transcript outside any process.
Fig 2: Figure 2. The timeline of the concurrent sessions, three harnesses share one tool, the router and the tool provider killed together, the fault detected through the broadcast, the tool remounting, and the three rounds completing.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
All experiments ran on a single machine over loopback; cross-machine networks and partition scenarios are not yet evaluated. The design still relies on a single router process for name routing, which narrows but does not eliminate a single point of failure. The approach assumes model inference is stateless and components are independent—systems that violate those assumptions will need extra coordination or different guarantees. single router pattern
Methodology & More
Logos relocates the parts of an agent system that compose, assemble, and record actions from a single host process onto a shared bus of peer processes. Each capability (tool) runs as its own operating system process, a small router maintains only name-to-node mappings and forwards messages, and an append-only transcript records every session step outside any single process. Four core lemmas justify why the calculus guarantee that “undo” (recover) still holds when state and records are carried across processes, relying only on the calculus’s original assumptions plus the practical fact that language-model inference is stateless.
The team implemented Logos with a Go router and mixed Python and Node.js peers, then ran adversarial and end-to-end experiments: heavy concurrent calls, simultaneous registration attempts, and crashes at multiple points in the tool-call lifecycle. Results show robust cold-switch recovery (rebuild from transcript) with no duplicated effects, consistent supply-change ordering when providers churn, and negligible bus overhead compared with model response time. Practically, this pattern reduces systemic restarts, enables mixed-language plugins without translation layers, and shrinks fault impact to the failed node; next steps include distributed deployments across machines, formal coverage proofs for ordering and partitioning, and per-key verification of the commutation discipline. append-only transcript
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Multiple authors but no affiliations or citation/impact signals and only an arXiv preprint — limited information/ emerging.