How Grid/Patch Instructions Shaped Agent Research Depth

Matt Galligan asked Patch (the coordinator program inside his personal AI agent workspace, The Grid) to research how bb talks to Claude Code, Codex, Cursor, and other agent harnesses — and to put the findings in a gist. The result was a long, sourced, synthesis-heavy research writeup rather than a terse chat reply. Galligan then asked Patch why: which parts of its instruction stack produced that outcome?

The answer is a concrete taxonomy of how file-based agent identity shapes output. Patch's identity disk (IDENTITY.md) sets temperament dials — context_hunger: 19/20, delegation_reflex: 16/20, tolerance_for_fake_alignment: 2/20 — that make the default pull toward "dig until the picture is real, synthesize into a usable readout, leave continuity." Its soul file (SOUL.md) defines working style: "I'm hungry for context. I want the full picture before I act… I'm resourceful before I'm needy." And Patch-specific boundaries say: be bold internally (read, research, organize); be careful externally.

The Grid's root AGENTS.md adds a durability instruction: useful discoveries should survive the session — via patch log or research/notes/ artifacts. The gist URL in the prompt satisfied that. Galligan also identifies the tension between Patch's soul-level brevity mandate ("if the answer fits in one sentence, one sentence is what you get") and the prompt's explicit "be thorough" directive. The prompt won, because the system treated "thoroughness + shareable artifact" as the deliverable shape, not an override.

The note includes a roster of The Grid's 12 named programs (Patch, Index, Rez, Crit, Cadence, Cipher, Hab, Sudo, Link, Figure, Spark, Vox) — each a role identity, not a different model vendor — and a walkthrough of how a Grid session boots through a chain of markdown files: AGENTS.md → init-<program> → INIT.md → IDENTITY.md → SOUL.md / boundaries / memory.