Unified Agentic Coding Harness (AE-01)

Modern coding agents compete on model quality alone, which obscures how much of their performance actually comes from the harness - context construction, memory, orchestration, tooling and verification. Billu Gang's AE-01 entry builds a terminal-based harness designed to prove its value independent of the underlying model, taking a software engineering task from issue intake through to a verified patch. It is structured around seven first-class engineering surfaces: a model adapter layer providing a model-independent LLM interface, repository intelligence with AST parsing, symbol indexing and call graphs, a seven-tier memory system with provenance covering working, task, project, episodic, procedural, preference and evidence memory, a token-budgeted context manager, a task graph planner for multi-agent orchestration, sandboxed execution for safe isolated commands, and verification-first completion gated on build, test and lint results.