Edward Y. Chang, Stanford University Position statement, v1.0. August 9, 2026.
The prevailing agent stack pairs chain-of-thought orchestration with a vector database for memory and hand-rolled retry logic for recovery. I take the position that this stack has reached its end as a foundation for long-horizon work, for reasons that are structural rather than incremental.
A chain of generated steps inherits four foundational limitations: agents validate their own plans; state lives in ephemeral context windows; no transactional safeguards exist, so there is no atomicity, rollback, or compensation; and interdependent constraints cross agent boundaries unowned. A vector database repairs none of these. It provides similarity recall and nothing more: no admission path for what becomes operational truth, no serializability under concurrent proposals, no compensation semantics, no notion of an effective present distinct from raw history, and no obligations that outlive a context window. Benchmarked on realistic multi-agent planning problems, standalone LLM pipelines produce fluent itineraries that violate interdependent constraints, and after a disruption they patch the affected step while silently breaking its dependents. Fluency without bookkeeping.
SagaLLM and ALAS mark the qualitative break: the chain of generated steps becomes a chain of admitted steps, with validation and state moved outside the generator. SagaLLM brings the saga discipline to planning: compensable units, persistent context, independent validation agents, and automated compensation, with the coordination plumbing itself generated by LLMs while transactional discipline constrains what that generated code may do. ALAS extends the guarantee to execution time, where rollback is too slow: a lightweight local compensation protocol repairs disruptions in place. In stochastic ride-sharing logistics it averaged 95.1 km against 118.9 km for baseline LLM pipelines, a twenty percent improvement at p below 0.01, and on the classical DMU job-shop instances extended with runtime disruptions it outperformed both standalone LLMs and traditional optimization solvers at maintaining operational continuity.
The field has run this movie before. Flat files with per-application locking gave way to the DBMS, which promoted recovery, isolation, and audit into substrate guarantees. The same move is now due for agents: admission, compensation, effective state, and durable obligations promoted from application hacks into substrate guarantees. Chain-of-thought survives as a within-step instrument; it ends as a foundation. Volume 3 of my trilogy completes the promotion with the Mnemosyne substrate and its ATP contract.
Full development: The Path to AGI, Vol. 1 (SagaLLM, ALAS) and Vol. 3 (Mnemosyne, ATP).