AI & Automation Briefing - July 26, 2026
OpenForgeRL Lets Teams Train Tool-Using Agents End-to-End With Open-Source RL Infrastructure
Training AI agents that use inference harnesses like Claude Code or Codex has been difficult because open-source reinforcement learning stacks cannot natively handle the stateful, multi-process nature of those harnesses. OpenForgeRL is a new open-source framework that addresses this by placing a lightweight proxy between the harness and the RL training stack, recording model calls as training data, and running each rollout in its own Kubernetes container. This architecture separates inference from training, allowing agents to be trained in the same harnesses and environments where they will actually be deployed. In benchmarks, the framework's tool-use model scored 31.7 pass^3 on ClawEval and its GUI agent scored 37.7 on OSWorld-Verified, with both outperforming open baselines of comparable size and matching larger models in GUI tasks. The research also found that RL training improves agentic behaviors like self-verification and multi-step planning, though error recovery remains a weak point across harnesses.
AREX Uses Recursive Self-Improvement to Build a Stronger Deep Research Agent
Researchers from BAAI published AREX, a deep research agent that improves itself recursively by alternating between two distinct phases: open-ended discovery and constraint-based verification. The asymmetry between these two phases is intentional. Discovery is broad and generative, while verification is strict and corrective. This loop allows the agent to refine its own outputs over successive iterations without human intervention. For automation builders, the pattern maps directly onto multi-step workflow designs where intermediate outputs are checked against defined criteria before passing to the next stage, similar to verification loops in n8n. Model weights are publicly available on Hugging Face and the system is accessible via a live application at arex-research.com.
Cognition Acquires Poke to Bring Conversational Personality to Devin Coding Agent
Cognition, the company behind the Devin coding agent, has acquired Poke, an AI assistant built around casual, friend-like conversation through SMS, iMessage, Telegram, and WhatsApp. The deal values Poke in the low nine figures. Cognition's stated goal is to make Devin feel more like a personable colleague than a software tool, borrowing Poke's interaction model and tone. Poke, launched in March 2026, generated over 100 million messages in its first three months across hundreds of thousands of users, but struggled with profitability due to high operating costs. Poke's most common use cases were productivity tasks including email management, reminders, and scheduling. The acquisition signals a broader strategic bet: that interaction design and agent personality will become a primary differentiator for AI products, not just an afterthought to model capability. For teams building client-facing agents or internal Slack bots, this is a concrete signal that how an AI communicates carries real product and retention value. Poke will continue operating on Apple's Messages for Business platform through at least the end of 2026.