Yathartha: A Protocol-Layer Treatment of Jagged Intelligence in Autonomous Agent Networks

Generative AI systems exhibit uneven performance across tasks in ways that do not map to human intuitions about difficulty. Dell'Acqua and colleagues named this the jagged frontier and measured its workplace consequence in a field experiment with BCG consultants: AI users produced output 40 percent higher in quality on tasks inside the frontier, and output 19 percentage points less likely to be correct on tasks outside it [2]. Mollick and Euchner later popularized the framing [1]. Gans formalized it as an information-economics adoption problem [3]. The literature is diagnostic and behavioral. It describes the phenomenon and prescribes organizational responses. It does not address jagged intelligence at the protocol layer. We present the Yathartha framework, a protocol-layer treatment of jagged intelligence with four contributions. First, we formalize the distinction between jaggedness (static unevenness in capability across tasks at a fixed time) and drift (change in behavior on a fixed task across time). These are orthogonal axes that the current agent-integrity literature conflates. Second, we define the CapabilitySurface primitive: a structured record of an agent's observed competence across discrete capability regions, published at baseline and refreshed on a declared cadence. Third, we propose three safety invariants that extend the Pratyahara NERVE specification [7] with capability-conditional drift detection, eliminating a specific class of false positives without weakening the underlying integrity guarantee. Fourth, we sketch empirical validation directions grounded in existing capability-evaluation benchmarks. ## v2 — changes from v1 Invariant **N-16 (Coverage-Conditional Drift)** is now stated in its operative action-level form, matching the TLA+ specification's enabling guard on `RaiseDrift`. v1 expressed N-16 as a state invariant on `driftFlags`, which TLC verification showed to be unsatisfiable (the set is append-only while `covered_regions` and `clockTick` are mutable). The action-level statement is what the design always required; v2 records that correction. Reference cross-citations to sibling protocols updated to concept (all-versions) DOIs.