Workday opens HR data to your agents over MCP, and Meta ships Muse Wonder
Workday opens its HR and finance data to your own agents over MCP, Meta ships a first frontier model, and BigQuery adds hard token quotas.
Three of today’s six items are about the same thing: giving agents access, and then working out who pays and who is allowed. Funny how the interesting work starts after the connection works.
1 · Top story
Workday lets your own agents in, with MCP and its permissions attached
At Workday Rising on October 13, Workday said its Trusted Agent Platform is now “headless”, so customers can build and run agents for HR and finance with their own tools and models. Agent-Ready Tools are MCP tools over Workday data, Live Data Query runs SQL against live data, and an Agent Harness hands each agent the user’s role, reporting line and policies, and logs each action. Five components are generally available, and Google ADK support is design-partner only.
The useful part for an architect is that permissions and audit stay inside Workday while the model and the orchestration stay yours, which is a reasonable split if you were dreading rebuilding HR access rules in a prompt. It’s Workday’s own release: no customer deployments are named and no pricing is public, so the first POC will tell you what the harness really costs in latency.
2 · Models & features
Meta’s first frontier model, Muse Wonder, is on the API
Meta announced Muse Wonder on October 12, available now through the Meta Model API and Muse Code. A larger Muse Wonder Full is limited to Muse Code subscribers. TechCrunch reports $4 in and $20 out per million tokens for Muse Wonder; Meta’s own page gives no price or context window.
Meta says it leads on agent, memory and coding benchmarks (it names Harvey’s Legal Agent Benchmark, FocusBench and SWE-Marathon) but publishes no scores in the post, so treat that as marketing until someone reruns it. Open-weight status isn’t stated either. It’s a fourth vendor to put behind your gateway for an A/B test, not a reason to move anything yet.
3 · Agents
DigitalOcean wants agents to improve from their own traces
On October 12, DigitalOcean announced Agent Learning Services (Insights, Signals, Evaluations and Simulations) in public preview, plus Agent Data Services in private preview. The data layer stores session traces, documents and vectors in one catalog with lineage, and the learning layer replays those sessions to test changes to an agent.
The pattern is the sensible one: captured production traces become your regression suite. The catch is that you’d be tying your trace store to one cloud’s runtime, and the private preview carries no SLA or compliance certifications, so it’s an idea to copy more than a product to adopt this quarter.
4 · Research
Your LLM judge may be the weakest part of the eval
A new preprint, revised October 14, builds a benchmark of stateful, tool-using customer-support workflows mined from real transcripts and runs 17 models against MCP-compatible mock environments. GPT-5.4 ranks first by a wide margin, and Qwen3.5-9B is the strongest open model under 20B. The sharper finding: five LLM judges agree on pass/fail only moderately (average kappa 0.52), so the ranking depends on who is judging.
If you gate releases on an LLM judge, that number should make you nervous. No official code or benchmark release is linked yet.
Try this: score 30 of your own transcripts with two different judge models and compute the kappa. Below roughly 0.6, add a rule-based check before you trust a trend.
5 · People worth following
Uber’s AI stack: one gateway, every harness
On The Pragmatic Engineer (October 14), Uber’s Uday Kiran Medisetty walks through the setup. A model gateway sits in front of Anthropic, OpenAI and Uber’s own open models, with a security layer on requests and responses, per-user spend tracking and audit logs. A wrapper called AIFX lets engineers move between Claude Code, Codex and OpenCode and keep their settings, and an MCP gateway reuses Uber’s existing Thrift and Protobuf registries. Uber says token costs fell about 50% over a few months.
The lesson I’d steal is the last one: the agent tooling was quick to build because a decade of developer tooling already existed. The gateway buys you freedom from lock-in at the price of one more layer to run.
6 · Architecture
BigQuery adds hard daily token quotas for its AI functions
Google announced on October 14 that cost management for BigQuery’s generative AI functions is generally available. Four daily quotas, per project and per user, cap input and output (including thinking) tokens for AI.GENERATE, AI.IF, AI.CLASSIFY and friends. Lowering a quota takes effect immediately, raising one needs a request, and embeddings aren’t covered.
The failure mode deserves attention. When a quota runs out mid-query, general-purpose functions like AI.GENERATE still succeed, but skipped rows get no value and carry a quota error in their status field. Your downstream job has to look for those empty rows. There are no per-query limits, and the per-user cap is per project.
Try this: run a batch under a quota set to about 10% of its needs, and check that your pipeline flags the empty rows. AI.COUNT_TOKENS gives a lower-bound estimate for sizing.
Nothing here is urgent. Pick the one thing that touches your own stack, and let the rest wait.
Researched and drafted with AI, checked against the linked sources. How we use AI →