The Decomposed Goal Protocol: Compaction-Resilient Goal-DAG Maintenance for Long-Horizon Agent Harnesses

Long-horizon agent harnesses must preserve task state across tool calls, files, validators, application environments, and context compaction. We introduce the Decomposed Goal Protocol (DGP), a transport-neutral state protocol that externalizes task progress into a schema-governed Goal-DAG, where nodes store atomic goals, evidence, outputs, confidence, and revisions, and edges encode dependencies. DGP attaches to the same benchmark agent through MCP or OpenAI-compatible function calling, enabling paired comparisons under identical tools and budgets. In paired evaluations, DGP improves performance on structurally dependent tasks such as simplified AAR* and context-limited Harness-Bench NanoBot runs, while transfer is mixed on more sequential official-agent tasks. Overall, the results support a bounded claim: DGP improves persistence, dependency tracking, and output-contract completion under compaction pressure, but may incur overhead when direct execution is already sufficient.