Murakkab: MIT + Microsoft publish a system that auto-optimizes agentic workflows

What changed: Researchers from MIT and Microsoft introduced Murakkab, a method that lets developers describe an agentic workflow in plain language and automatically selects models, tools, and hardware configurations — then adjusts those choices at runtime to trade off speed, cost, or energy use. The team reports marked reductions in compute and energy compared with hand-configured deployments.

Why it matters: Founders and engineers running multi-step agent systems can stop treating orchestration as fixed code and instead use declarative intent to get better resource efficiency; that lowers operating cost and carbon footprint and makes long-running agents easier to maintain.

Try/watch: Pilot Murakkab-style optimization on one non-critical workflow (e.g., a nightly ETL + analysis pipeline) to measure token/compute savings before applying to production-critical agents; watch for changes in latency and failure modes when the runtime swaps model variants.

AutoLabs: PNNL’s agent translates experiment goals into robot actions

What changed: Pacific Northwest National Laboratory released AutoLabs, an agentic system that converts high-level experimental goals into robot-specific instructions for automated lab equipment (tested on the Big Kahuna robot) and published results and code the same day. The paper and a GitHub release aim to make automated experimentation faster and more accessible.

Why it matters: Lab managers, instrument vendors, and startups building automated R&D pipelines can accelerate throughput (the team reports 5–10x more experiments) by lowering the scientist–engineer handoff; this is practical automation rather than a toy demo.

Try/watch: If you run lab automation, download the code and run it in a sandboxed environment to validate safety checks and edge-case handling; monitor reproducibility and whether the agent requires human-in-the-loop gates for hazardous steps.

Elsevier adds agentic features to LeapSpace for research workflows

What changed: Elsevier expanded its LeapSpace research workspace with agentic capabilities — features like a Writing Coach, Claim Radar, and Compare Tables — that operate over Elsevier’s indexed full-text literature and metadata, and go live for customers immediately. The launch emphasizes traceable citations and researcher approval for any changes.

Why it matters: Universities, CROs, and enterprise R&D teams can embed agents that do literature synthesis while preserving traceability and citation provenance, which reduces the manual work in review, drafting, and evidence-checking. Buyers should evaluate how provenance and access controls are implemented before adopting.

Try/watch: Test LeapSpace agent outputs against domain experts for citation accuracy and hallucination rates; require configurable approval gates for any manuscript edits or claim changes.

RingCentral AIR Pro: agentic CX moves from suggestions to actions

What changed: RingCentral upgraded AIR Pro to include native AI agents that can autonomously handle customer interactions, start outbound outreach (appointment reminders, payment notices), and perform intelligent handoffs that transfer full context to humans. The feature set was announced at CCW Las Vegas and is positioned for enterprise contact centers.

Why it matters: Contact-center operators and outsourcers can reduce human handling time and automate routine flows, but they must update compliance, monitoring, and escalation rules because agents will be executing actions on behalf of customers.

Try/watch: Run a controlled pilot on low-risk interactions (notifications, confirmations) with full logging and human override; monitor error rates, compliance with payment/consent rules, and customer satisfaction metrics.