AI Agents

A multi-case study reads three open coding-agent harnesses at pinned commits — LangChain's deepagents, Earendil's pi, and DeepSeek's dsh — and traces their commit histories. Despite opposing starting philosophies (batteries-included vs. minimalism vs. plugin-everything), the mature two converge on five elements: a commoditized loop, an append-only replayable session record, model quirks stored as data, progressive context disclosure, and explicit extension seams. The held-out third exhibits all five, in one case reusing another's code. The one dimension nobody built: external, tamper-evident verifiability.

Most RCA evaluations score whether an agent localizes the responsible microservice. This work argues that endpoint correctness hides whether the agent actually reconstructed the fault-propagation path an on-call SRE needs to act. Across 3,500 diagnostic trajectories on a public benchmark, it finds agents that name the right service yet can't trace the propagation. Failures come from omitting decisive evidence, misreading retrieved telemetry, or substituting unsupported inference. The resulting DiagGuard architecture surveys observations before localizing, then audits the diagnosis against them.

ARQ refines C/C++ CodeQL queries using execution-grounded evidence rather than labeled data or templates. The insight: synthesize a program, run it, and when its execution disagrees with the query's verdict, you've located an FN (genuinely vulnerable, query silent) or FP (safe, query fires) weakness. An LLM repair loop treats those disagreements as ground truth. Refining 12 official queries with three commercial models, ARQ detects up to 119.8% more true positives at ≥98.0% precision, and fixes long-standing issues in existing queries.

AUSO treats skills as having a lifecycle: first learnable knowledge, then capability-building, then invoked only when they improve a specific decision. Existing methods either keep skills external, fully internalize them, or switch objectives on noisy task-level success rates — assigning uniform importance to every action in a trajectory. AUSO instead computes an action-level signal by evaluating each sampled action under skill-conditioned and skill-free contexts, coupling that with trajectory-outcome advantage so beneficial skill-sensitive actions get stronger updates and harmful ones are suppressed. Reports gains on three benchmarks.

TRACE is a deployed catalog-enrichment agent: a ScoutAgent triangulates multimodal evidence across merchant catalogs, syndicated feeds, and identity-matched web search to propose attribute values with supporting evidence, while a JudgeAgent verifies each against that evidence and decides to publish or route to human review. Offline, proposals hit 98.2% accuracy at 74.7% coverage; in production across four verticals it raised impression-weighted enrichment coverage 90.4%, and surfacing enriched attributes lifted checkout conversion 0.48%. Domain-specific, but the propose/verify split with evidence grounding generalizes.