acceptodds
Desk rejected at ICLR 2027

From Playbooks to Rulebooks: Extending Agentic Context Engineering (ACE) for Episodic Workflow Learning

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Tool-using large language model agents often repeat workflow failures because prior execution experience is not converted into reusable guidance. This paper introduces a novel application of Agentic Context Engineering (ACE) to a single, stable agentic workflow task, in which deliberately simulated tool-level barrier scenarios are paired with human-crafted ground-truth resolutions to teach an agent how to recover from recurring operational exceptions. Through the Generator-Reflector-Curator training loop, ACE compares failed or incomplete trajectories with expected human-defined outcomes and distills corrective guidance into an episodic rulebook: a persistent, inspectable memory artifact containing generalizable exception-handling and recovery strategies for future executions. We evaluate this approach in a controlled, multi-step notification workflow containing happy-path, no-notification, and simulated barrier scenarios, without modifying model weights. The rulebook was learned from 30 training examples and evaluated on a disjoint, randomly selected set of 50 held-out examples from a 196-example test bed. The ACE-trained Generator correctly identified and notified the required recipients in 41 of 50 cases (82%), compared with 24 of 50 cases (48%) for an untrained baseline using the same Generator, tools, mock data environment, and initial Context Playbook. The framework may extend to complex, repeatable industry workflows in which agents must coordinate multiple tools, apply business rules, and recover safely from incomplete or ambiguous information. By converting operational exceptions, human corrections, and approved recovery paths into inspectable guidance, ACE could support more consistent, auditable, and context-aware agent behavior within established governance and human-oversight controls.

Primary area:
reinforcement learning
Keywords:
Agentic Context Engineering, reinforcement learning, agentic memory, episodic memory, procedural memory, ACE, tool-augmented memory, workflow agents, context engineering, human feedback, rulebooks, playbooks

Related papers

Discussion (0)

Sign in to comment.