Sanctuary System Research Institute

Independent research on agent context substrates and graded-oversight systems for autonomous AI work.

The current program studies one question: what structure does a tool-using agent need its context to have, and what does it cost when the prevailing substrate withholds that structure? The findings so far are stated in four claims, each carrying its receipt.

Interactive: experience the substrate difference — the same 50 skills loaded flat versus descended as a tree, with the context cost counted live.

The four claims

  1. Agent operation is graph-shaped; the prevailing substrate is flat. An agent in operation is a flow through context-states: nodes are states, edges are admissible transitions, and the space the agent moves through is precisely the structure of its context. A flat substrate withholds that structure, so the agent must reconstruct it at every step. Token counts are the symptom, not the problem: fifty tools loaded upfront is not a token problem but a structure problem. Reified far enough, the structure is not a tool the agent consults but what the agent runs as — stated as the program's organizing conjecture, not a proven result. Optional corroboration, kept deliberately out of the foundation: mechanistic work on in-context learning identifies edge-shaped attention machinery (induction heads; Olsson, Elhage et al. 2022), and models are separately documented to degrade when relevant content is buried among irrelevant content ("Lost in the Middle"). Both are treated as corroboration of the structural claim, never as its basis. Receipt: Paper 1, §1 and Figure 1; the demotion of the mechanistic point is explicit in §1.1.
  2. MCP tools and agent skills shipped flat; the fix is a tree on either side. A flat tool surface (every schema loaded upfront) and a flat skill library (a directory of self-contained packages) are the same bug. A tree descended on demand loads O(depth) rather than O(#tools); the "MCP versus skills" token debate compares two flat representations and asks which is cheaper, which is the wrong axis — the answer on both sides is to impose a tree. Receipt: Paper 1, §4 and Figure 1, calibrated against the platform's own published measurement (150,000 tokens upfront vs. 2,000 on demand).
  3. "Progressive disclosure" does not apply to itself. Skills disclose progressively within a package — frontmatter, body, referenced files — but a skill cannot cause a deeper skill to be injected, so disclosure stops at depth one and a library of skills remains a flat forest exactly at the scale where disclosure was the point. The correction is the missing level itself: a coordinate-addressed, navigable, self-maintaining tree over the skill library. Receipt: github.com/sancovp/skilltree (MIT, v0.2.0, 62 tests passing); Paper 1, §6, §8, and Appendix A, including the runtime load-semantics observations.
  4. This is tractable now, and was not before. Hierarchical agents that select a minimal toolkit are no longer exotic to build. Two conditions delayed the correction — models were not previously reliable enough at instruction-following and tool use to drive their own navigation of a tree, and the tools were built flat, which is a default, not a law. Both conditions have lifted; the remaining distance is inexpensive to close. Receipt: Paper 1, §10; running code at skilltree and graph-skills-experimental.

The frame

No intent is claimed anywhere in this program, and none is required — the shape of the failure is identical whether it was chosen or merely shipped quickly. These tools are infrastructure: a population of builders stands on the platform's defaults, so a default that mis-shapes context is not one team's cost but a levy paid across every project built on top, continuously. The program closes on a demand rather than a complaint: a single tree that self-manages, and an agent that can navigate it.

Publications