KDCube
← AI Industry News

Agent Harnesses Become the Control Plane as Meta Joins the Frontier

Workday and DXC are turning the agent harness into the governed runtime for enterprise work, routing every agent through MCP tools, central traces, and audit controls. Meanwhile Meta shipped its first frontier models, Muse Wonder and Muse Wonder Full, and New Relic detailed always-on LLM-as-a-judge evaluation ...

Lead signals

Highlights

  • Meta Superintelligence Labs shipped Muse Wonder and Muse Wonder Full on Oct 12 — its first frontier-class models, built explicitly for agentic work with "Active Memory" and "Active Focus," though only the smaller tier is exposed via API. (Meta)
  • Workday made Sana its front door and took the Agent Harness generally available, so every agent — Workday's, a customer's, or a third party's — runs through one governed runtime with memory, guardrails, regional controls, and audit traces. (Workday)
  • DXC put LangChain's LangGraph and LangSmith inside mission-critical IT operations, licensing the stack to build, evaluate, and govern its production OASIS agents. (DXC)
  • New Relic detailed AI Evaluation, an async "LLM-as-a-judge" service that scores live telemetry for hallucinations, prompt injection, and data leaks — public preview due in November. (New Relic)

Key Signals

  1. Meta enters the frontier — and gates the strong tier - Oct 12

    Muse Wonder Full lands near Claude Sonnet 5.5 and GPT-6 Astra on the Artificial Analysis Intelligence Index, but Meta ships only the weaker Muse Wonder through the Model API; the full model sits behind a Muse Code subscription. (VentureBeat) For builders this is another agent-native frontier option (Wonder at $4/$20 per million tokens) — with the caveat that benchmark-topping capability isn't the thing you can actually call programmatically. (Meta)

  2. The "harness" is becoming the governance unit - Oct 13

    Workday's Trusted Agent Platform now offers headless access to HR and finance data, exposing "Agent-Ready Tools" as MCP endpoints so any MCP-speaking agent gets an approved path to read and act. (Workday) DXC's parallel move — standardizing OASIS on LangGraph/LangSmith, claiming a 95% cut in resolution time across 500 test cases — shows the same pattern: a single runtime that orchestrates, evaluates, and logs every agent action rather than per-app bots. (DXC)

  3. Evaluation shifts from pre-launch to always-on - Oct 6

    New Relic's AI Evaluation scans production telemetry continuously instead of sampling isolated LLM calls, tying agent behavior to transaction-level business impact — a direct answer to its own stat that one in four enterprises runs agents in production with no monitoring at all. (New Relic)

Why It Matters / What To Watch

  1. Pick runtimes by governance surface, not demo quality.
    • Workday, DXC, and the broader "harness" pattern converge on MCP + centralized traces as the control plane — evaluate platforms on whether tool access, audit, and regional data controls are first-class. (Workday)
    • Watch how Workday's MCP "Agent-Ready Tools" authorize third-party agents; headless data access is powerful and a fresh attack surface, echoing this month's MCP trust-boundary concerns. (Workday)
  2. Treat frontier model access as a procurement constraint.
    • Meta's strong tier is subscription-gated; confirm which model you can actually call via API before building on benchmark numbers. (VentureBeat)
    • If you're committing to Muse Wonder's agent features (Active Memory, Active Focus), validate cross-session continuity against your own eval suite rather than vendor framing. (Meta)
  3. Stand up continuous evals before scaling agents.
    • New Relic's LLM-as-a-judge approach lands in preview next month; teams running unmonitored agents should plan for telemetry-based scoring of injection and leakage now. (New Relic)

Quick Links