Skip to content

Stateful Iteration State-Carry: Typed Persistent State for Long Agent Loops

Carry agent state across iterations through a state-read tool instead of replaying the full transcript each turn — converts O(n²) loop token cost to O(n).

Learn it hands-on: Remember, Don't Re-Read — guided lesson with quizzes.

When this pattern applies

This refactor pays back only under specific conditions. Apply it when all of the following hold (Jabbarvaziri, 2026):

  • The loop is long-horizon: tens of iterations per run, not single-digit conversational turns.
  • Per-iteration observations are large: source code, benchmark output, full search results — not single scalars.
  • The loop runs unattended in production: the engineering cost of typed state and a checkpointer is amortized across many runs.
  • The agent's next decision usually needs only a subset of prior state, not the whole trajectory.

If any condition is missing — short loops, tiny observations, exploratory single-shot runs — the simpler stateless approach with prompt caching at the provider tier hits the same cost line for less engineering work.

The cost curve the pattern targets

A stateless ReAct loop appends each Thought-Action-Observation triple to the message history, then re-sends the whole transcript on the next call. Per-call input grows linearly with step n, and total cost across N steps is O(N²) — every prior observation is re-billed on every subsequent inference. The pattern is mechanical and reproducible across providers (Augment Code, 2026; Towards Data Science, 2026).

Stateful state-carry lifts the experimental record out of the transcript into a typed object that lives outside the prompt. The agent reads specific fields via a tool only when the current decision needs them. The conversation window stays approximately fixed-size; total cost across N steps becomes O(N) (Jabbarvaziri, 2026).

The paper measured both regimes on two benchmarks:

Task Iterations Stateless tokens Stateful tokens Reduction
Hyperparameter tuning (small observations) 15 24,465 2,492 90%
Code optimization (large observations) 40 1,275K 627K 52%

Optimization quality was comparable on both tasks — token reduction did not degrade outcomes (Jabbarvaziri, 2026).

How to apply it tool-agnostically

The paper's reference implementation uses LangGraph, but the pattern is framework-agnostic — state lives outside the prompt, accessed by tool call. It resembles a code-native memory substrate specialized for iteration loops:

  1. Define a typed state object for the loop's experimental record — current best metric, last hyperparameter set, recent failure traces, working files. Keep fields minimal; every additional field becomes another schema migration (Mem0, 2026).
  2. Expose state through tools the agent can call: read_state(field), update_state(field, value), list_recent_attempts(n). The agent invokes these when it needs context, instead of expecting full history in the transcript.
  3. Persist state via a checkpointer, not in-memory. InMemorySaver loses everything on restart and undermines durability (Towards AI, 2026); production loops use Redis, Postgres, or DynamoDB-backed stores (AWS, 2026).
  4. Trim the message window to recent turns only. The state object is the source of truth for prior iterations, not the transcript.

Why it works

The causal mechanism is purely about where state lives, not about model reasoning. A stateless loop encodes the experimental record in the message transcript. The inference call must re-process that transcript every turn, and the provider re-bills the full prefix on each call. A stateful loop encodes the same record in a typed object outside the prompt and exposes it through tools. Per-call input is bounded by the working set the current step touches, not by cumulative history. The asymptotic effect (O(N²) → O(N)) is a direct consequence of decoupling the experimental record from the transcript (Jabbarvaziri, 2026). The 90% and 52% reductions are the gap between a quadratic and a linear cost curve at finite iteration counts, not optimizations on top of the same architecture.

When this backfires

Conditions under which the pattern is worse than the alternative:

  • Short loops with stable prefixes. Below ~10 iterations, prompt caching at the provider tier already converts the dominant cost line to roughly O(1) for the static portion. Anthropic charges ~10% of input price on cache hit, OpenAI ~50% (NeuralTrust, 2026). The stateful refactor adds engineering cost the cached stateless loop avoids.
  • State schema churn. When the experimental shape changes often — new fields, renamed metrics, restructured observations — every schema change risks breaking persisted checkpoints. Projects have failed because "every additional field increases complexity exponentially" and the state object became a monolith (Mem0, 2026).
  • Concurrent or multi-replica execution without isolation. Shared state corrupts silently under concurrent writes; the failure surfaces several state transitions downstream from the cause, making root-cause hard (Focused.io, 2026).
  • Tasks that need the full trajectory. Causal debugging, exploratory branching, transcript replay — pruning to typed state discards the audit trail that made the loop debuggable.
  • In-memory persistence in restart-prone environments. InMemorySaver loses everything on restart; durable state requires a Postgres/DynamoDB backend, shifting cost from tokens to operations (AWS, 2026).

Contrast with prompt caching

The two approaches attack the same cost curve from different layers:

Prompt caching Stateful state-carry
Where the fix lives Provider infra Application code
Engineering cost Order static content first; lock tool schemas Design typed state, write state-read tools, operate a checkpointer
What gets cheaper Static prefix (system prompt, tool defs) Growing observation history
Best for Stable prefix, short to medium loops Long loops, large observations
Failure mode Cache misses on prefix mutation State schema churn, concurrent corruption

They are complementary at the boundary case: a long-horizon stateful loop still benefits from a cached static prefix on the residual transcript. For short loops with stable observations, the two approaches compete.

Key Takeaways

  • The token saving comes from where state lives, not from changing the model — typed state outside the prompt makes per-iteration input independent of step count.
  • Empirical savings scale with loop length and observation size: 90% at 15 iterations with small observations, 52% at 40 iterations with large observations.
  • The refactor only pays back for long unattended loops with large observations; short or exploratory loops should reach for prompt caching first.
  • State schema discipline matters as much as the lift itself — every added field becomes a migration surface, and shared state under concurrency corrupts silently.
Emerging — single recent source Last reviewed: 2026-06-16