Agent Operations Commit Wholly or Roll Back Wholly

The runtime platform guarantees that every agent operation either commits all of its state changes or reverts the runtime's in-memory state to the pre-operation snapshot. Partial completion of an agent operation is not a representable outcome in the runtime's persistent state; the platform carries the all-or-nothing guarantee, so agent code does not have to track and undo its own intermediate side effects when something fails mid-operation.

Why it is held

Agent operations are characterized by failure modes that traditional application code is not engineered for. An LLM call mid-operation may return garbage; a tool call may error; a precondition may be violated by a concurrent agent; an external service may time out. Without a platform-level atomic guarantee, every agent operation that touches more than one object would have to carry hand-written compensating-action logic for each step that might fail -- the same pattern each team rebuilds, with bugs. The platform inverts that burden: the runtime's transactional envelope owns the rollback, and agent code can be written as if every operation either succeeds in full or did not happen at all.

The atomic guarantee operates at the in-memory state graph level, not at a separate database or persistence layer. When the runtime rolls back an atomic call tree, the rollback covers object references, accumulated state mutations across all objects touched in the call tree, queued callouts and events, and per-task resource accounting. This is materially different from "the database transaction was rolled back" -- there, transactional rollback covers rows in a persistence layer while in-memory caches and application-level state may already be inconsistent. The runtime's atomicity reverts the runtime itself; there is no in-memory residue of the failed operation to reconcile after the fact.

The platform-level position matters because partial state is the structural failure mode in a system where multiple agents share state. When agent A's operation halts halfway through (some side effects landed, some didn't), agent B reading state cannot distinguish "A is in progress" from "A finished" from "A failed mid-step." The all-or-nothing guarantee removes this entire class of confusion: any state agent B sees is either a state agent A's operation produced in full or a state untouched by agent A. Without that guarantee, every multi-agent interaction inherits the partial-state ambiguity, and each team rebuilds coordination scaffolding to recover from it.

This Conviction is held because atomicity treated as a primitive -- every line of agent code can rely on it without engineering it -- is what platform adoption purchases. Patterns get rebuilt by each team, badly; primitives are rebuilt once, by the runtime, and become invisible infrastructure that lifts the application's complexity ceiling. Trading the primitive away to broaden the platform's apparent applicability would surrender the load-bearing claim the project rests on.

What it asks

Honoring this Conviction asks the project to defend the atomic envelope as a first-class boundary, not as an internal implementation detail.

Drift recognition

The Conviction has weakened when the project's outputs treat partial state as a real possibility worth handling, rather than as an impossibility the platform prevents.

Sources

Relations