autoguide

AutoGuide — context-aware guidelines

Re-implementation of AutoGuide (Fu et al., 2024) on daedalus's harness. Accumulation runs each training task with ReAct+Reflexion (up to 3 retries) to collect (success, failure) pairs; for each pair (at most 3 per task) it finds the step where the trajectories diverge, summarizes the shared prefix into a context, and extracts one guideline for it (Algorithm 1). The memory is a context → guidelines dictionary. At test time, every turn, the agent's context is identified, matched against the bank's contexts, and up to k = 3 of the matched context's guidelines are injected into that turn's prompt; when no context matches, nothing is injected (Algorithm 2).

Run

# 1. build the guideline bank from the training split
uv run python -m references.autoguide.accumulation --config references/autoguide/configs/tau2_accumulation.yaml
# 2. evaluate on the test split (5 runs)
uv run python -m references.autoguide.run --config references/autoguide/configs/tau2_inference.yaml

Configs ship for AppWorld (appworld_*), τ²-bench retail (tau2_accumulation, tau2_inference) and AutomationBench Operations (automationbench_*). run takes --bank and -k overrides on top of the common flags. Outputs: outputs/baselines/autoguide/memory/<name>/ (pool.json = the bank, pool.log.json = one row per guideline, experiences/, task_summaries/, traces/) and outputs/baselines/autoguide/inference/<name>/ (run_<i>/, evaluation.json, retrieval/ = the per-turn context decisions and their cost).

autoguide:
  bank_path: "outputs/baselines/autoguide/memory/tau2/pool.json"
  k: 3                          # guidelines injected per turn
  context_model: null           # context identification/matching/selection; null = solver model
  extraction_model: "gpt-5.4"   # guideline extraction (Eq. 2)
  extraction_reasoning_effort: "high"
  max_retries: 3                # Reflexion retries when gathering experience
  reflection_model: "gpt-5.4"
  max_pairs_per_task: 3

Inference cost: one context-identification call per turn, one matching call per context string not yet seen by the worker, and one selection call when the matched context holds more than k guidelines. Totals are printed at the end of a run and recorded in retrieval/.

Paper → code

paper	here
Algorithm 1 (bank construction), deviation step	accumulation.py (build_bank, deviation_index), over references/common/experience.py
context identification / matching / guideline extraction / selection	modules.py + prompts/context_*.txt, prompts/guideline_*.txt
Algorithm 2 (test time)	retrieval.py::ContextAwareRetriever, installed by agents/<benchmark>.py
the guideline dictionary	guidelines.py::GuidelineBank, stored as a daedalus pool

Deviations from upstream

• The ALFWorld prompt variants are used (the closest to a tool-calling environment), with "a text-based ALFRED task" reworded as "an interactive task".
• Context identification runs zero-shot: the ALFWorld examples teach a household vocabulary that would mis-key every context here. The extraction prompt keeps the paper's two format examples.
• Two inputs the figures abbreviate are filled explicitly: the identified context reaches the extraction prompt as SUMMARIZATION: <context>, and the matching prompt receives the new context under Newly Generated Summarization:. {Task description} is one line naming the benchmark and domain (prompts/task_description.txt).
• k is a parameter (the paper's figures hardcode 2 or 3); k = 3 is the best value in the paper's own ablation.
• Reflexion's reflection prompt is daedalus's post-attempt reflection prompt (AutoGuide ships none).
• On τ²-bench, context identification, matching and selection run on gpt-5.4, the auxiliary model of every method in the paper (context_model: gpt-5.4); AutoGuide runs them on the agent model, as the AppWorld and AutomationBench configs do.
• Not implemented: the WebArena extraction variant (against human demonstrations) and AutoGuide + Reflexion at test time.