LogAct: Ensuring Agent Reliability via Shared Logs

LogAct proposes a new abstraction that decomposes agents into state machines executing based on shared logs, addressing the reliability challenges of agent deployment in production environments (asynchrony, failure recovery, behavior audit). It enables auditable, interceptable, and recoverable actions, providing a solid guarantee for the production deployment of agents.

Large language model-driven agents have capabilities like autonomous planning and tool calling, but production deployment faces three key challenges:

1. Asynchrony: The timing and results of interactions with multiple external services are hard to predict;
2. Failure recovery: It’s difficult to restore to the correct state when the agent or environment fails;
3. Behavior audit: The decision-making process is opaque, making problem tracing challenging. Existing solutions mostly focus on capability enhancement, with insufficient research on reliability assurance.

LogAct decomposes agents into state machines centered around shared logs, drawing on the event sourcing pattern and optimizing it. Key attributes include:

1. Pre-execution visibility: Actions are written to logs before execution, facilitating review and intervention;
2. Pluggable interception mechanism: Actions are reviewed via independent voters;
3. Consistent failure recovery: Replay/rollback from logs to a consistent state. Architecture components include a shared log layer (persistent action records), state machine engine (drives state changes), voter framework (extensible review), and recovery manager (failure recovery).

LogAct leverages LLM reasoning to analyze execution history, enabling:

1. Semantic recovery: Understand failure semantics and adopt targeted strategies (retry, alternative solutions, etc.);
2. Self-debugging: Review execution traces to identify inefficient patterns or error sources;
3. Token usage optimization: Reduce redundant interactions in multi-agent clusters to save computing resources.

Experiments verify LogAct’s effectiveness:

1. Failure recovery: Efficiently restore to a consistent state in various failure scenarios; recovery time depends on log size;
2. Performance overhead: Acceptable latency in normal paths with no unpredictable peaks;
3. Security interception: Successfully block all unwanted actions, with only a 3% drop in availability of benign functions;
4. Multi-agent optimization: Reduce redundant interactions by approximately 25% and save resources.

LogAct emphasizes auditability as a fundamental attribute to meet regulatory compliance and troubleshooting needs; it combines distributed system patterns (event sourcing, CQRS) with LLM to deeply customize agent features; its pluggable architecture supports custom governance rules. As agents take on key business roles, reliability infrastructure like LogAct will become indispensable.

Current limitations:

1. Mainly focuses on single-agent reliability; multi-agent collaboration scenarios need further exploration;
2. The voter framework may become a performance bottleneck in high-throughput scenarios. Future directions: Combine formal verification to provide stricter guarantees; expand support for complex action types (creative decisions, fuzzy boundary operations).