SCLATE: month-long agent calendars on a hybrid clock, not a custom loop per harness
Sep 26 Jung/Bilkhu et al. (Apple): open event scheduler + hybrid simulated clock compresses 30-workday MetaClaw to median 5.9h (162×); ten harness×memory configs × ten models; post-train through unmodified Claude Code/Hermes/Codex.
Source: SCLATE: A Substrate for Continual-Learning Agent Training and Evaluation
Youngmok Jung, Sirajul Salekin, Henry Tran, Javier Movellan, Zhao Huang, and Manjot Bilkhu (Apple) target a missing substrate for continual-learning agents—systems of model + harness + memory that must interleave tasks with session stop/start, crons, and memory consolidation across multi-day horizons. Existing suites either hardwire a built-in assistant or schedule only the benchmark’s events, so every harness–benchmark pair reinvents a scheduling loop. SCLATE makes both sides adapters on one open event scheduler. A hybrid simulated clock flows with wall time while the agent (or a background job) is busy, then jumps idle gaps—compressing overnight and weekend silence into pure execution time via libfaketime so unmodified Claude Code / Hermes / Codex crons and memory sidecars see the scenario timeline without source edits. An in-container proxy records tokens and log-probs for every model call, attributed to harness vs memory, turning the same runs into post-training rollouts.
On seven ported benchmarks (MetaClaw, GAIA-2, AppWorld, LongMemEval, PersonaMem, SWE-Gym, SWE-bench Verified), a 30-workday MetaClaw arc spanning 40 calendar days finishes in a median 5.9 real hours (~162× speed-up; 142× over all runs combined). Across 100 MetaClaw configs the scheduler dispatches 77,700 benchmark-side and 12,400 agent-side events; GAIA-2 fires 13,050/13,050 environment events with 98.7% at the scenario-specified delay. Head-to-head: ten harness×memory configurations × ten models show no single memory wins for every model—added Mem0/Hindsight/GBrain underperforms native memory in 11/50 Claude Code / Hermes cells—and models differ sharply in harness use (Gemini-3.7F spends 7.6% of tool calls on session-search vs ≤1.6% for others). Post-training Qwen3.5-4B through unmodified harnesses lifts held-out MetaClaw by up to +11.8 pp (Claude Code + Mem0) and SWE-bench Verified Pass@1 9.8%→26.5% under Claude Code, with 6.8× fewer file lines read on matched tasks.
Why it matters for code review
Always-on factories that only measure single-session SWE patches miss the overnight shape of real agent work: session boundaries, nightly consolidation, mid-turn notifications. SCLATE’s useful receipt is infrastructural—put crons and memory on the same clock as the workload, compress idle time without lying about event order, and score model–harness–memory together. Continuous verified momentum compounds when you can replay a month of factory life in hours and reject “we bolted on Mem0” claims that don’t beat native memory for that model. FOMO for more overnight sessions without a shared scheduler just multiplies one-off harness scripts.