Slipstream: Trajectory-Grounded Compaction Validation for Long-Horizon Agents

Researchers introduce Slipstream, a system that validates LLM agent trajectory compression by running compaction asynchronously alongside continued agent execution, enabling independent validation of summarized context. The approach improves task accuracy by up to 8.8 percentage points while reducing latency by 39.7% on long-horizon coding and web-browsing tasks.

Slipstream addresses a fundamental problem in long-horizon LLM agent systems: trajectory compaction, which summarizes accumulated context to manage token limits, currently lacks reliable validation mechanisms. Traditional synchronous compaction forces agents to resume from summaries that may omit critical information, yet validation criteria cannot exist until after the summary influences downstream behavior. The innovation lies in parallelizing compaction with continued agent execution on original context, generating both a candidate summary and the agent's next reasoning steps from the same pre-compaction state. This asynchronous approach creates an independent validation signal by comparing whether the compressed summary preserves the agent's forward intent and key constraints.

This work emerges from a critical bottleneck in scaling LLM agents beyond token limits. As agents tackle complex tasks like software engineering (SWE-bench) and web navigation, accumulated trajectory grows exponentially, necessitating compression. Current methods risk silent failures where agents produce coherent but incorrect behavior based on corrupted summaries. Slipstream's trajectory-grounded validation represents a meaningful architectural shift that prioritizes correctness over raw efficiency.

The improvements—8.8 percentage points accuracy gain with 39.7% latency reduction—suggest the technique delivers both safety and performance gains. This impacts developers building production agent systems, particularly in domains requiring high reliability. The reduction in end-to-end latency while improving accuracy indicates the asynchronous approach eliminates wasted recomputation on invalid paths. Developers deploying long-horizon agents should evaluate whether Slipstream's validation framework applies to their architectures, as silent failures from trajectory corruption represent a critical risk in autonomous systems.

• →Asynchronous compaction validation independently checks trajectory summaries against actual agent behavior, eliminating validation gaps in synchronous approaches.
• →Slipstream achieves up to 8.8 percentage point accuracy improvements on long-horizon tasks while reducing latency by nearly 40 percent.
• →The system validates both forward intent preservation and critical facts retention, preventing silent failures in autonomous agent execution.
• →Independent validation signals generated from pre-compaction state eliminate the structural problem of validating summaries agents rely upon.
• →Results span diverse domains including coding tasks and web browsing, suggesting broad applicability to production LLM agent deployments.