Failure Modes in Multi-Turn Agents
Trajectory 1: Customer Service Agent
An agent handling shoe size exchanges ingests unindexed, conflicting legacy documentation (30-day cutoff), poisoning its reasoning and wrongly rejecting an eligible request under the 60-day policy.
Trajectory 2: Customer Service Agent
Multi-hop logistics queries for regional inventory stock and freight dispatch return large, unpruned payloads, causing explosive context growth and saturating the token budget.
Agentic Search & Retrieval
Trajectory 1: FreshCart Coding Agent
The agent attempts to locate exact code constants using pure vector semantic search and policy concepts using keyword search, failing to adapt or recover without hybrid retrieval.
Trajectory 2: FreshCart Coding Agent
Fixed character-count splitting without language-aware AST parsing splits model validation logic across boundaries, leading to hallucinated data constraints.
Context Formatting & Compaction
Trajectory 1: Cinema Research Agent
Compacting summaries of summaries across turns progressively strips vital clues (jazz score, European jury award) needed to identify an obscure animated film.
Trajectory 2: Cinema Research Agent
Externalizing full document bodies to a scratchpad without returning the external document ID pointer strands retrieved records outside the agent's reach.
Agent Memory & State Persistence
Trajectory 1: Doctor's Appointment Assistant
Demonstrates memory writing pipelines reconciling newly paraphrased patient preferences against long-term vector memory to prevent duplicate rule entries.
Trajectory 2: Doctor's Appointment Assistant
A stateless assistant using exact flat-file keyword search fails to equate "Friday" with "end-of-week", forgetting user preferences across sessions.