Agentic Mission Orchestration for Agricultural Robots: Source Code and Support Data

Abstract

The following material supports the article Agentic Mission Orchestration for Agricultural Robots: Coupling Agronomic Validity, Edge AI Reliability, and Data Governance, submitted to Precision Agriculture (Springer Nature). This deposit corresponds to the first revision (R1) of the manuscript and supersedes the material deposited for the original submission. The archive holds 487 files. Contents agents/: 10 modules, comprising the orchestrator, the five specialist agents, the monolithic single-prompt baseline, the deterministic rule engine, and the resource-matched baseline conditions. prompts/: the 11 system prompts used by the agents and by every baseline condition. data/missions/: 23 mission specifications, covering M1 to M6, the held-out missions H1 to H4, and the test missions T1 to T10; earlier variants of T1 to T3 are retained under a .v1 suffix. data/agronomic_kb.yaml and data/capability_registry.yaml: the agronomic knowledge base and the capability registry. eval/: the scoring code, the ground truth, and the per-run records, aggregated scores, and summaries for every reported condition. Five entry points at the archive root: run_orchestration.py, run_monolithic.py, run_rule_engine.py, run_baseline.py, and run_b1_matched_pilot.py. Where each reported result comes from Section 5.1, conditions and models → agents/, prompts/, run_orchestration.py, run_monolithic.py Section 5.2, mission scenarios → data/missions/ Sections 5.3 and 5.4, metrics and component-wise scoring → eval/score_v3.py, which contains the rule-based semantic-equivalence table Table 4 → eval/results/v2/eval_summary_v3_component_M1M4.md Section 5.5, deterministic and hybrid baseline → agents/rule_engine.py, eval/results/rule_engine/, eval/results/rule_engine_t/ Section 5.6 and Table 5, agent-order sensitivity → eval/results/b5_permutation/records/, 10 run records Section 5.7 and Table 6, held-out evaluation → data/missions/H1 to H4, eval/results/held_out/, 32 run records Section 5.8 and Table 7, compute-matched baselines → agents/baseline_conditions.py, eval/results/b1_same_model/, eval/results/b1_baselines/ Section 6, Tables 8 to 11, and Figure 3 → eval/results/v2/, eval/results/frontier/, eval/results/smaller/ Ground-truth audit → eval/ground_truth.yaml Each run record carries a _meta block naming the model identifier used for that run and the observed client-side latency. What is not included Raw provider-response bodies and separately persisted retry and parsing-failure events are not part of this deposit, because the original runs did not record them. Observed latency includes API time and YAML-parse retry time, but individual retry events, provider queue time, and monetary cost were not persisted separately. These limits are stated explicitly in the manuscript rather than implied here. Files are under restricted access during peer review.