The Pulse
Apple Researchers Build SCLATE to Train Long-Running Agents
Apple researchers introduce SCLATE, a system for training and evaluating AI agents across long, multi-session tasks. Its simulated clock compresses a 40-day scenario into a median of 5.9 hours, while experiments test agent memory and post-t
AI.info Team · Written with AI ·
“A hybrid simulated clock runs these events on a shared timeline, flowing in real time while the agent works and skipping idle gaps, which compresses a month-long scenario into hours.”
Youngmok Jung, Sirajul Salekin, Henry Tran, Javier Movellan, Zhao Huang and Manjot Bilkhu, authors of “SCLATE: a Substrate for Continual-Learning Agent Training and Evaluation”
Apple researchers propose SCLATE, an execution system for training and testing AI agents that work across multiple sessions. The paper, submitted to arXiv on September 26 and marked as submitted to ICLR 2027, targets a scheduling problem: long-running agents have their own events—such as session restarts, scheduled tasks and memory updates—that standard benchmarks do not coordinate with benchmark events.
Rather than build a custom scheduler for each benchmark-agent pairing, SCLATE gives benchmarks and agents adapters to register events on one shared timeline. The system also captures model calls through an in-container proxy, recording tokens and log probabilities so evaluation runs can feed into training. The authors describe it as a way to evaluate agents in their existing harnesses and memory systems, without modifying those components.
A shared clock for benchmarks and agents
SCLATE’s clock advances in real time while an agent is working or a background task is active. When nothing is running and future events are scheduled, it skips the idle interval to the next event. That lets a scenario retain its sequence and timing while avoiding the real-world wait between sessions or scheduled updates.
In the paper’s MetaClaw test, a 30-workday scenario spanning 40 calendar days completed in a median of 5.9 hours across 100 runs. The reported median speed-up was 162 times; across all runs, the speed-up was 142 times. The authors also report that SCLATE fired all 13,050 scheduled GAIA-2 environment events, with 98.7% of checkable events delivered at the scenario’s specified delay.
Seven benchmarks, ten models
The researchers ported seven public benchmarks covering personal assistance, software engineering and multi-session memory. They compared ten combinations of agent harnesses and memory systems across ten models. Results show that memory is not a simple add-on: while each tested memory configuration beat its harness’s no-memory baseline in the MetaClaw comparison, no single memory system performed best for every model.
Performance also varied by task. On GAIA-2, scenarios requiring agents to react to events during a run remained difficult across the tested model-and-harness combinations. The paper reports maximum pass rates of 37.5% for its Adaptability category and 34.4% for Time. The authors say the shared scheduler makes these asynchronous scenarios easier to run consistently, but the results also expose how far agents still struggle with events that arrive after a task begins.
Post-training changes how the model uses memory
The team used SCLATE to post-train Qwen3.5-4B through unmodified agent harnesses. In the software-engineering experiments, the trained model’s SWE-bench Verified Pass@1 rose from 9.8% to 26.5% under Claude Code, a gain of 16.7 percentage points. Under Codex, it increased from 28.8% to 33.6%. On 42 matched tasks under Claude Code, the trained model read 6.8 times fewer file lines and hit its context limit on 31.0% of tasks, down from 97.6%.
On held-out MetaClaw days, accuracy increased by 7.6 to 11.8 percentage points in five of six configurations; three of those gains were statistically significant, according to the paper. The trained model also wrote memory records with more procedural rules and more detail about its own actions. Those results are specific to the tested models, tasks and configurations; they do not establish that the same gains will transfer to other agents or workloads.
The hard part is more than elapsed time
SCLATE’s central contribution is experimental infrastructure: it coordinates benchmark events with the internal routines that let agents continue work across sessions, then captures model calls for post-training. Its experiments show why that distinction matters. Memory systems and harnesses interact differently with different models, and improved performance in one evaluation does not settle which setup will work best elsewhere.
For now, the paper reports a research prototype and results on seven benchmarks, rather than a broadly validated training recipe. Its clearest measured demonstration is the 40-calendar-day MetaClaw arc, completed in a median 5.9 hours while the scheduler preserved the scheduled event timeline.