The Idempotency Illusion: Why AI Agents Need Stable Intent Over Generated Hashes

Posted on Oct 21 • Originally published at tamiz.pro

The Idempotency Illusion: Why AI Agents Need Stable Intent Over Generated Hashes

In distributed agent architectures, idempotency is the safety net that prevents a single user request from spawning duplicate actions. Yet a growing pattern—using freshly generated UUIDs or cryptographic hashes as the sole key for idempotency checks—creates a silent failure mode: the idempotency illusion. This article unpacks why generated hashes break idempotency, what stable intent replaces them with, and how to wire your agents for true repeatability.

Idempotency seems straightforward: repeat the same input, get the same result. In HTTP APIs, this maps to retry-safe headers and deduplication keys. In AI agent frameworks, the contract extends to tool calls, state mutations, and cross-turn continuations. The assumption is that if two requests are "the same," the system can recognize them and suppress the duplicate.

Many agent platforms generate a fresh hash—often a UUIDv4 or a SHA-256 of the prompt—for every inbound request. The intent is to use this value as a cache key for "already processed?" lookups. But several structural flaws make this approach brittle:

Stable intent shifts the key from a generated artifact to a canonical representation of what the user actually wants. Instead of hash(prompt), intent is derived from the user’s goal, normalized against a stable schema. This can take several forms:

To replace hash-based idempotency, integrate the pattern at three layers:

import hashlib, json, redis. asyncio as redis def normalize_intent ( raw: str ) -> str: # Strip noise, lowercase, sort keys, extract canonical form cleaned = ‘ ’. join ( raw. lower (). split ()) return hashlib. sha256 ( cleaned. encode ()). hexdigest () async def idempotent_call ( redis_conn: redis. Redirect, intent: str, coro ): key = ‘ intent: ’ + intent cached = await redis_conn. get ( key ) if cached: return json. loads ( cached ) result = await coro () await redis_conn. set ( key, json. dumps ( result ), ex = 3600 ) return result Enter fullscreen mode Exit fullscreen mode Pitfalls and Edge Cases Over-normalization: Aggressive normalization can merge distinct intents. Always include a “user context” fragment (e.g., account ID, session tier) in the intent key.

Partial failures: If the agent crashes after executing the action but before caching the intent, the next retry will duplicate the action. Use two-phase commits or idempotency tokens returned to the caller. Multi-step agents: For agents that chain calls, each step must propagate the same stable intent, or the deduplication window shrinks to a single hop.

Adversarial inputs: Malicious prompts designed to produce the same normalized intent but different outcomes require policy-level allowlists, not just hash checks. When to Stick with Generated Hashes Generated hashes are not obsolete. They excel as transient session identifiers, trace correlation keys, or “poison-pill” markers for dead-letter queues.

Use them when the cost of a duplicate is low and the input space is bounded. Pair them with a secondary intent check for safety-critical paths.

Q: Can I use both generated hashes and stable intent together? A: Yes. Treat the hash as a fast-path cache key and the stable intent as the authoritative deduplication source. Collisions in the hash layer are caught by the intent layer. Q: What if my agent framework does not expose structured outputs? A: Normalize the raw text on the ingestion side. A preprocessing middleware that runs deterministic normalization before the model sees the prompt is a low-effort entry point. Q: How do I choose a persistence layer for the intent registry? A: If your agents are stateless serverless functions, use a durable KV store with TTL (DynamoDB, Redis with Expiry). For long-running agent processes, an in-memory map with checkpoint persistence works fine.

The core problem with using generated hashes is that they are often derived rather than intrinsic. If your agent generates a hash based on the current state of its internal reasoning trace, that hash changes every time the agent "thinks" differently, even if the ultimate intent remains identical. To solve this, we must shift the idempotency key from a derivative of the process to a canonical representation of the intent.

Consider a user request: "Refactor the payment_service to use Stripe instead of Braintree."

We need a layer that extracts the "semantic core" of a task before executing it. This core becomes the idempotency key.

Before generating a key, normalize all inputs. This means:

import json import hashlib from typing import Any, Dict class IntentNormalizer: def __init__ ( self, excluded_fields: set = None ): self. excluded_fields = excluded_fields or { " timestamp ", " request_id ", " session_token " } def normalize ( self, payload: Dict [ str, Any ]) -> bytes: """ Creates a canonical JSON representation of the payload excluding ephemeral fields. """ # 1. Filter out ephemeral fields clean_payload = { k: v for k, v in payload. items () if k not in self. excluded_fields } # 2.

Sort keys recursively for deterministic serialization clean_payload = self. _sort_keys ( clean_payload ) # 3. Serialize with sorted keys and compact separators canonical_json = json. dumps ( clean_payload, sort_keys = True, separators = ( ', ', ': ' )) return canonical_json. encode ( ' utf-8 ' ) def _sort_keys ( self, obj: Any ) -> Any: if isinstance ( obj, dict ): return { k: self. _sort_keys ( obj [ k ]) for k in sorted ( obj )} if isinstance ( obj, list ): # Note: Sorting lists can be tricky if order matters. # Only sort if the field is defined as a "set" in your schema. return [ self. _sort_keys ( item ) for item in obj ] return obj def generate_stable_intent_key ( payload: Dict [ str, Any ]) -> str: normalizer = IntentNormalizer () canonical_bytes = normalizer. normalize ( payload ) return hashlib. sha256 ( canonical_bytes ). hexdigest () Enter fullscreen mode Exit fullscreen mode Step 2: The Execution Gate Now, wrap your agent’s execution loop with this gate. The agent should always check the store before doing expensive work (LLM calls, tool invocations).

class IdempotentAgentExecutor: def __init__ ( self, kv_store, agent_logic ): self. kv_store = kv_store self. agent_logic = agent_logic def execute ( self, task_payload: Dict [ str, Any ]): # 1. Generate the stable key idempotency_key = f " agent:run: { generate_stable_intent_key ( task_payload ) } " # 2. Check if already completed cached_result = self. kv_store. get ( idempotency_key ) if cached_result: # Log that we're reusing a result return { " status ": " replayed ", " result ": cached_result, " idempotency_key ": idempotency_key } # 3.

Lock to prevent concurrent execution of the same intent # Use a distributed lock (e.g., Redis SETNX with timeout) if not self. _acquire_lock ( idempotency_key ): return { " status ": " in_progress ", " idempotency_key ": idempotency_key } try: # 4. Execute the actual agent logic result = self. agent_logic. run ( task_payload ) # 5. Persist the result with a TTL self. kv_store. set ( idempotency_key, result, ex = 3600 * 24 ) # 24h TTL return { " status ": " success ", " result ": result, " idempotency_key ": idempotency_key } except Exception as e: # Do NOT persist errors as successful results raise e finally: self. _release_lock ( idempotency_key ) def _acquire_lock ( self, key: str ) -> bool: # Simplified for illustration; use a proper distributed lock library in production return self. kv_store. set ( f " lock: { key } ", " locked ", nx = True, ex = 300 ) def _release_lock ( self, key: str ): self. kv_store. delete ( f " lock: { key } " ) Enter fullscreen mode Exit fullscreen mode Handling Ambiguity: When Is an Intent "The Same"?

This is the hardest part of the "Idempotency Illusion." Two requests look the same but might have different underlying contexts.

If your task_payload only contains {"action": "summarize"}, both generate the same hash. This is a false positive.

You must include immutable references to the context in the hash, not the content itself.

Update your IntentNormalizer to include context_refs:

payload = { " action ": " summarize ", " context_refs ": { " article_id ": " abc123 ", " model_version ": " gpt-4o-2024-01-01 " # Model version matters for determinism } } # This will produce different keys for Article A vs B, even if the instruction is identical. Enter fullscreen mode Exit fullscreen mode Testing for Stability Unit tests for idempotency are straightforward. You just assert that the key generation is deterministic across varying non-semantic inputs.

import pytest def test_intent_hash_is_stable_to_noise (): payload1 = { " task ": " calculate ", " inputs ": { " a ": 10, " b ": 20 }, " request_id ": " req-123 ", " timestamp ": " 2023-10-27T10:00:00Z " } payload2 = { " task ": " calculate ", " inputs ": { " b ": 20, " a ": 10 }, # Different order, same semantic " request_id ": " req-456 ", " timestamp ": " 2023-10-27T11:00:00Z " } key1 = generate_stable_intent_key ( payload1 ) key2 = generate_stable_intent_key ( payload2 ) assert key1 == key2, " Hashes should match despite noise and key ordering " def test_intent_hash_changes_with_semantics (): payload1 = { " task ": " calculate ", " inputs ": { " a ": 10, " b ": 20 }} payload2 = { " task ": " calculate ", " inputs ": { " a ": 10, " b ": 21 }} key1 = generate_stable_intent_key ( payload1 ) key2 = generate_stable_intent_key ( payload2 ) assert key1!= key2, " Hashes should differ for different inputs " Enter fullscreen mode Exit fullscreen mode Conclusion The "Idempotency Illusion" stems from treating AI agents like stateless functions where hash(input) = output. In reality, agents are stateful, non-deterministic systems that interact with dynamic environments.

By shifting from generated hashes to stable intents, you turn retries from a source of chaos into a reliable safety net, allowing your agent systems to operate with the confidence of traditional, deterministic infrastructure.