<\/p>\n
Errors compound over long trajectories, and nondeterminism (sampling, tool variability) complicates reproducible evaluation and makes it difficult to attribute failures to specific components (Wang et al., 2022). When outcomes deviate from the plan, verifier and replanning loops (reflection, tree-style search over alternatives) help recover by updating beliefs from new sensor observations and re-issuing skill-level commands (Yao et al., 2023a; Shinn et al., 2023). Privacy constraints often limit logging and debugging, which makes it harder to diagnose failures and to build high-quality evaluation sets, and evaluation is expensive without labeled multimodal ground truth.<\/p>\n<\/p>\n
Comparability requires fixing retrieval policies and reporting ablations over context budgets and planner depth; otherwise improvements may reflect system tuning rather than underlying capability. This design reduces context load, makes outputs evidence-backed via citations, and supports targeted debugging via intermediate artifacts and traces (Yao et al., 2023b). Division of labor across sub-agents (segmenter, retriever, summarizer, verifier) mirrors modular tool-using architectures, and verification can include timestamp citation checks and stability checks via reruns (Karpas et al., 2022; Schick et al., 2023; Yao et al., 2023a; Shinn et al., 2023; Wang et al., 2022). Safety requirements often prohibit open-ended exploration, and robots operate under hard constraints (collision avoidance, force limits, workspace boundaries) that must be enforced regardless of language intent.<\/p>\n<\/p>\n