Long-Horizon Dependability Beyond Single Sessions

Determine how the Claude Code architecture—comprising the agentic query loop, context construction and compaction, subagent delegation, and session persistence—can continue to support long-horizon dependability when autonomous work spans beyond a single session.

Background

The architecture centers on turn, session, and subagent units. As autonomous work extends to multi-session or multi-day programs (e.g., research workflows), it is unknown whether current mechanisms suffice or require new coordination primitives or memory substrates.

This question ties together earlier concerns about compaction limits, subagent isolation, and append-only persistence, now examined at horizon scales.

References

How the architecture documented in \Cref{sec:arch,sec:turn,sec:context,sec:subagent,sec:persist} (whose primary units are the turn, the session, and the sub-agent) continues to support long-horizon dependability as autonomous work extends beyond a single session is an open question.

— Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems  (2604.14228 - Liu et al., 14 Apr 2026) in Section 12.4 (Horizon Scaling: From Session to Scientific Program)

This would allow us to test whether covert strategies remain internally consistent across memory boundaries and handoffs, whether goal-task conflicts compound when strategic behavior is distributed, and whether patterns observed here, such as the role of instrumental goals or the reasoning-action gap.

— SchemeArena: Factorized Stress Testing of Scheming in LLM Agents  (2609.08126 - Ruan et al., 8 Sep 2026) in Section 7, “Further Discussion,” subsection “Persistent and Multi-Agent Settings”

Whether a human-AI collaboration can sustain those obligations over the multi-year lifetime of a solver, rather than over a single task, remains an open question.

— Aether.jl : A High-Performance 3D MHD and Multifluid Dust Code Written in a Dynamic Language with an Interactive Human-AI Development Framework  (2608.14048 - Ho, 14 Aug 2026) in Section 5, Human-AI Development Framework, opening paragraph

Our 16K results suggest that combining such mechanisms with learned proactive compaction could further improve these systems, which we leave for future work.

— AutoCompact: Learning When to Compact Context in Long-Horizon Coding Agents  (2610.02163 - Zhang et al., 1 Oct 2026) in Limitations and future work, Conclusion

Open research questions. (i)~What constitutes an agent's footprint checkpoint: can we checkpoint its context at an intermediate reasoning step, so that recovery need not restart the task? (ii)~How can we extend existing recovery mechanisms to deal with heavy non-determinism? (iii)~How can durability acknowledgments from heterogeneous systems compose into a single, trustworthy commit?

— Tracking State Footprints: How Agents Can Transact  (2610.03140 - Mraz et al., 2 Oct 2026) in Section 5.3, “Durability and Fault Tolerance”