[arXiv Research] Agentic Workflow Improvements — 2026-10-05 #65820

Paste one of these as a new issue or comment to kick off implementation with @copilot (one prompt per opportunity, in the same order as above):

@copilot Implement: Extend gh-aw workflows with an optional final deterministic verification step that evaluates the assembled result and cross-stage invariants after all jobs or sub-agents complete, independently of each stage’s local success. The verifier should consume declared handoff outputs and emit a typed pass/fail result with actionable mismatch details through the existing safe-output/audit path. in gh-aw's Composed-workflow verification component. Rationale: FinSkillBench found that curated resources improved individual-task performance, yet its nine-stage probe produced a near-optimal result with zero mandate adherence even though every stage reported success. gh-aw’s compiler and typed safe-output layer can make global verification explicit: compile declared workflow invariants into a terminal check over the actual handoff artifacts, preventing locally valid outputs from silently composing into a globally invalid result. Source: Knowledge or Calculator? Decomposing the Skill Premium in Verifiable Financial Agent Workflows.

@copilot Implement: Extend gh-aw's agent-drain pipeline with a deterministic preprocessor that converts long execution traces into typed diagnostic events, marks signals such as failed tool calls, retries, malformed outputs, and repeated observations, and serializes the most causally useful events under a configurable token budget for one rubric-guided diagnostic model call. in gh-aw's Trajectory diagnostics and token optimization component. Rationale: LiteTrajEval's offline rule profiles, heuristic signal marking, fixed-budget serialization, and single structured judge map directly to gh-aw's agentdrain log-mining and token-optimization components. The paper reports improved failure-localization alignment while reducing evaluation cost by about 6× and time by more than 8×, making this a concrete route to cheaper, repeatable workflow diagnostics without changing runtime engines. Source: Lightweight, Rubric-Guided Trajectory Evaluation for Production AI Agents.

@copilot Implement: Extend gh-aw workflow configuration so authors can assign a high-capability engine to infrequent strategy/planning steps and a lower-cost engine to implementation/refinement steps under a shared cost or token budget. Compile shared, invariant context into a stable prompt prefix and move step-specific feedback into a suffix, then record best-so-far task score against cumulative usage so workflows can compare routing policies by quality gained per unit cost. in gh-aw's cost-aware engine routing component. Rationale: FrugalEvo's core mechanism maps to gh-aw's pluggable engine integration and compiler-generated prompts: the compiler can preserve role-specific engine choices and arrange prompt segments to improve prefix reuse. Its budget-aware evaluation also fits gh-aw's token-optimization telemetry, giving workflow authors a concrete measure for deciding whether heterogeneous routing improves operational efficiency rather than merely reducing calls. Source: FrugalEvo: Towards Cost-Aware LLM-Guided Program Evolution.