• The automation ceiling is structural, not technical — first formal theory identifies three grounds where human participation persists regardless of AI capability
• Agent training enters the harness era — OpenForgeRL enables end-to-end RL inside production agent frameworks, closing the train-deploy gap
• LLMs can't track evolving intent — all model families fail when user goals change mid-conversation, a capability invisible to static benchmarks
METHOD: Rather than asking "how far can automation extend?", the authors invert the question: "where do its conceptual limits lie?" They identify three distinct grounds for persistent human participation — technical/complementarity (capabilities AI lacks), normative/developmental (participation is valuable for human agency), and emergence (objectives that are co-defined through the interaction itself and cannot be pre-specified).
KEY RESULT: The "emergence grounds" — the most novel contribution — show that for activities where goals emerge through interaction (design, therapy, negotiation, strategy), human participation is constitutive of the target being produced, not a workaround for imperfect AI. Human–AI co-construction is a permanent structural feature, not a temporary transition state.
🎓 SO WHAT: Banking is the canonical domain for all three grounds. Regulatory AI — normative grounds: accountability cannot be delegated. Advisory/complex lending — emergence grounds: client needs surface through conversation. Risk modeling — complementarity: human judgment fills gaps AI can't anticipate. This paper provides the intellectual framework for resisting full-automation pressure from vendors and consultants.
METHOD: Modern agents (Claude Code, Codex, OpenClaw) rely on complex harnesses for tool use and multi-turn reasoning — but these harnesses make end-to-end RL training nearly impossible. OpenForgeRL solves this with a lightweight proxy that intercepts model calls + a Kubernetes orchestrator that isolates each training rollout, decoupling training from inference so agents learn inside the real harness.
KEY RESULT: Using only hundreds to a few thousand tasks, OpenForgeClaw hits 31.7 pass³ / 55.9 pass@3 on ClawEval, while OpenForgeGUI reaches 37.7 on OSWorld-Verified and 63.0 on Online-Mind2Web — matching or surpassing models several times larger. RL improves self-verification, tool coverage, and multi-step planning. But error recovery remains stubbornly weak — the next frontier.
🎓 SO WHAT: This is the missing infrastructure for production agent deployment. Today's banking agent pipelines train on static datasets then deploy into harnesses they've never seen. OpenForgeRL closes that gap — train in the same harness you deploy in. Harness architecture IS a strategic choice, not an implementation detail.
METHOD: A framework that transforms static single-turn tasks into multi-turn conversations where user intent incrementally reveals, revises, and redirects itself. The original evaluation protocol is preserved — no new annotation needed, enabling existing benchmarks to serve as controlled testbeds for intent-tracking capability.
KEY RESULT: Across all tested model families, strong static performance does NOT transfer to evolving-intent settings. The drop is substantial and consistent. The capability to faithfully track and act on evolving user intent — essential for genuine collaborative agents — is systematically absent, yet completely invisible to static benchmarks.
🎓 SO WHAT: This directly explains why enterprise AI deployments underwhelm. Your bank's chatbot passes static tests then fails in production because real customers revise their requests mid-conversation. The evaluation gap IS the deployment gap. We're benchmarking AI on static tasks then deploying it into dynamic co-construction environments.