CoAgent brings serializable concurrency control to multi-agent LLM systems
Researchers introduce CoAgent, a toolcall middleware implementing the MTPO protocol that achieves serializable multi-agent execution within 5% of serial correctness at a 1.4× speedup, outperforming classical 2PL and OCC approaches on contended workloads.
Score breakdown
CoAgent replaces the abort-and-retry waste of OCC and the blocking delays of 2PL with an advisory protocol that lets LLM agents self-repair conflicts, achieving serializable correctness while preserving meaningful concurrency gains that classical mechanisms cannot sustain.
- 01Authors: Hongtao Lyu, Dingyan Zhang, and Mingyu Wu (published on ArXiv, 2026-06-13)
- 02Classical 2PL and OCC fail for LLM agents because single-agent transactions span minutes of inference and writes take effect immediately
- 03The MTPO (Monotonic Trajectory Pre-Order) protocol fixes a serialization order at launch and applies writes speculatively in place
Hongtao Lyu, Dingyan Zhang, and Mingyu Wu identify a fundamental mismatch between classical concurrency control mechanisms and the realities of multi-agent LLM systems. When multiple agents — coding agents, devops agents, document agents — operate in parallel against shared state, they enter the same conflict territory that database concurrency theory has studied for decades. However, classical approaches break down: a single agent transaction spans minutes of inference, read sets are broad and opaque rather than statically inferable, and the live state agents act on admits neither fork nor buffer, meaning writes take effect the moment they execute. Locks block long inference intervals, and OCC's abort-and-retry strategy discards minutes of work on every conflict.
This shifts control from the runtime to the agent itself in an advisory model.
The paper's key insight is that LLM agents possess a capability classical transactions lack: the model inside each agent can judge whether a conflicting write actually invalidates its plan, and can repair only the operations that depended on it. This shifts control from the runtime to the agent itself in an advisory model. The resulting protocol, MTPO (Monotonic Trajectory Pre-Order), fixes a serialization order at launch, serves each read the order-filtered value, and applies writes speculatively in place. When a conflict occurs, a one-way notification asks the affected agent to re-judge and patch its plan, while the framework mechanically undoes and reorders misplaced writes using saga-style inverse operations that each tool registers in advance. At quiescence, the execution is serializable in the pre-decided order.
CoAgent realizes MTPO as toolcall middleware, with a privileged component called ToolSmith that grows a library of footprint-declared, undoable tools online. Evaluated on ten contended workloads, CoAgent stays within 5% of serial correctness at a 1.4× speedup and near-serial token cost — a regime where 2PL and OCC surrender nearly all concurrency gains. On a bash-only target system, ToolSmith grows a 25-tool library online and lifts the task pass rate from 45/71 to 63/71 at 0.80× the time and 0.86× the cost.
Key facts
- 01Authors: Hongtao Lyu, Dingyan Zhang, and Mingyu Wu (published on ArXiv, 2026-06-13)
- 02Classical 2PL and OCC fail for LLM agents because single-agent transactions span minutes of inference and writes take effect immediately
- 03The MTPO (Monotonic Trajectory Pre-Order) protocol fixes a serialization order at launch and applies writes speculatively in place
- 04Agents self-repair conflicting plans via one-way notifications, with the framework mechanically undoing misplaced writes using saga-style inverse operations
- 05CoAgent is implemented as toolcall middleware; its ToolSmith component grows footprint-declared, undoable tools online
- 06On ten contended workloads, CoAgent stays within 5% of serial correctness at a 1.4× speedup and near-serial token cost
- 07On a bash-only target system, CoAgent raises task pass rate from 45/71 to 63/71 at 0.80× the time and 0.86× the cost
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 16, 2026 · 23:11 UTC. How this works →