Hey there 👋 For a few weeks the agent news has circled one worry: once an agent is running, how do you keep it from doing something dumb. Inboxes to approve its work. Kill switches. Identity brokers. A second agent watching the first. All of it kicks in after the agent is already loose in your systems. Today's news sits one step earlier, and I think it is the more useful place to stand. Not how you keep the agent in line once it runs, but what the thing even is. Whose weights is it running. Whose name is it acting under. Who wrote the skill you plugged in last week, and did anyone sign off. Provenance is the boring word for knowing what you are running before you run it. My bias, for what it is worth: the expensive agent failures this year will come from not knowing what was running, long before anyone reaches for a kill switch. So today is about the quieter half of the problem.
The Big ThingAbacus.AI put a self-hostable agent model on Hugging FaceAbacus.AI released three open-weight models on September 10 built for agent work, and the one worth your attention is Smaug Agentic. It is a 2 trillion parameter model based on Kimi K3, tuned for the long, multi-step loops that make most chat models fall apart after a few turns. Smaug Flash targets always-on agents that need to answer fast, and Smaug Mini is a 27B model for multimodal and lighter reasoning jobs. All three are downloadable on Hugging Face today, or callable through Abacus's RouteLLM API if you would rather not host them yourself. The download is the story. For a year the practical choice for a capable agent model was to send every call out to a frontier vendor and trust that your data handling held up on the way. Smaug Agentic runs inside your own cloud VPC or on an in-house GPU cluster, which means the prompts, the tool outputs, and whatever customer data moves through a long agent run never leave your walls. For a regulated team that has been watching the agent wave from the sidelines, that data-stays-home property matters more than any benchmark number attached to it. Grade the benchmark carefully, because Abacus supplied it. The company says its fine-tuning adds 15 to 20 percent on long-running agent loops at no extra cost, and that the line runs 10 to 100 times cheaper than closed frontier models. Those are Abacus's own figures, measured Abacus's way, with no independent agent benchmark to check them against yet. Treat the mechanism as real; treat the multiplier as a marketing number until someone outside the company runs the test. Ship it? Deploy now, as a test. Pull Smaug Agentic from Hugging Face, stand it up next to whatever model your agents run on today, and put the same real workflow through both for a week. You are measuring two things: whether the task quality holds, and what your token bill and your data exposure look like once the model lives on your side of the wall. If it holds, the self-host option just got a lot more affordable than it was in August. Source: Abacus.AI (abacus.ai/smaug) and PRNewswire, Sep 10 2026; weights on Hugging Face. Benchmark and cost figures are Abacus's own.
Tour de Headlines💳 Three payment rivals agreed to identify agents the same way. Ant International, Mastercard, and Visa said on September 10 they will connect their separate agent-identity protocols, Visa's Trusted Agent Protocol, Mastercard's Verifiable Intent, and Ant's Agentic Mobile Protocol, into one Know Your Agent interoperability layer, wired together through BuildFin.ai and convened alongside the Monetary Authority of Singapore. Plenty of vendors have already made payments a checkpoint for agents. What is new here is cooperation: three networks that compete on everything agreed on a shared way to onboard and recognize a paying agent across all of them, and that shared identity layer is the missing piece for agentic checkout at any real scale. Temper it, because there is no technical spec, governance body, or timeline yet, and the 3 to 5 trillion dollars of agent commerce they forecast for 2030 is a projection worth discounting until a real spec exists. If you are building anything that lets an agent spend, design your agent-identity and consent tokens now so you are ready when the spec lands. 🔌 Nutanix shipped a control point that governs what your agents are allowed to touch. Enterprise AI 2.8 went generally available with the MCP Gateway in its Agent Gateway now GA, a single place to manage how software agents connect to your applications and data. It adds fine-grained identity and access, custom roles, model-sharing controls, and support for air-gapped NVIDIA NIM, so a regulated shop can run production agents without opening a path off its own network. NKP 2.19 rounds it out with a built-in AI catalog and CNCF Certified Kubernetes AI Conformance. The provenance angle: every agent connection routes through one governed control point you can watch and revoke, instead of each agent wiring its own way into your systems. It reached GA in late August, so it is available now for Nutanix shops. If you run one, stand up the MCP Gateway first, because an agent you cannot see connecting is an agent you cannot govern. 🏦 A fund-admin vendor shipped the auditable version of an agent workflow. Fund Recs launched an Agentic Platform and a managed AI Ops service on September 11, putting specialized agents (support, document extraction, template building, resolution and controls) inside an oversight layer built on MCP, with three agents already in production. Two design choices are worth stealing whatever business you are in. First, client data never leaves the Fund Recs environment and is not used to train any model, which is how you sell automation to a compliance team. Second, the agents prepare the work and a human signs off before anything becomes a control action, with deterministic rules running the final step. That is the shape of agentic automation a regulated buyer will sign off on: the agent drafts, a person authorizes, the machine executes the checked rule. Copy the pattern even if you never touch fund administration.
Sponsor Your team's calls are full of signal. Measure it. RapportScore reads the calls your team already records and scores how your people communicate, then turns it into coaching they can use. Stop guessing whether a rep is building trust or talking past the buyer, and see it in the numbers. See your team’s score → |
Tool of the Day🔦 The provenance pass (the pattern, not a product)A one-afternoon audit of what your agents really are, before something they do forces the question. Today's stories all point at one weekend job, and you can do it with tools you already own. Walk your live agents and answer four questions for each. One, whose model is it running, and could you name the exact weights and where they execute. Two, what identity does it act under, a shared human login or its own scoped credential. Three, what outside skills and MCP servers did you plug in, and did anyone read them before you granted keys. Four, when a tool call fails, does the agent surface the failure or quietly report success. Write the answers in a plain table. The gaps are your to-do list. Most teams have never run this pass, which is why the first agent incident tends to be a surprise about something that had been running the whole time. Do it once now and you turn provenance from a postmortem question into a checklist. See a governed setup →
Worth a Click- Smaug on Hugging Face. The open-weight agent models from today's lead. Pull Smaug Agentic, run it beside your current stack, and grade Abacus's 15-to-20-percent and 10-to-100-times-cheaper claims against your own workload. (abacus.ai)
- Ant, Mastercard, and Visa on Know Your Agent. The framework announcement in the networks' own words. Read it for the shape of where agent identity is heading, and note how much is still principle rather than spec. (businesswire.com)
- Fund Recs Agentic Platform. The regulated-finance blueprint in full: MCP-based agents that prepare the work, a human who signs off the result, and client data that never leaves the environment or trains a model. (funds-europe.com)
This was a provenance week, quietly. Strip the logos off everything that shipped and one question sits under all of it: can you say, for certain, what agent is running. Abacus let you own the model outright, weights on your own hardware, so you know exactly what it is. Ant, Mastercard, and Visa started building a way to prove which agent is spending your customer's money. Nutanix put every agent connection through one governed control point you can watch. Fund Recs showed the regulated-finance version, where the agent drafts and a person signs the result. Last month the worry was how to control an agent after it runs. This week the more useful worry moved earlier, to knowing whose model, whose identity, and whose connection you are trusting before it runs. You can finally own the whole thing top to bottom, which is exactly why proving what it is became the job. See you tomorrow.
— Ron |
