Your agent spends while away

Hey there 👋

I spent twenty minutes this morning reading a pricing breakdown of Google's new Argon model, hunting for the catch. It sits about halfway down the page. The two-dollar input price is a promo, the real rate is double, and one hands-off task burns somewhere near sixty thousand output tokens. The model is not the trap. The invoice is.

That is the shape of the whole issue. Google opened its long-horizon model to everyone. Workday wrapped a governance harness around its agents. DXC pushed LangGraph into regulated server rooms. Rocket put agents on the mainframe. Four companies, one quiet agreement: the hard part of shipping an agent moved off the model and onto the scaffolding you build around it.

- Ron

The Big Thing

Gemini 4 Argon just opened to everyone, and it keeps working while you sleep

Following up on the gated rollout we flagged on October 3: Google moved Gemini 4 Argon out of its trusted-tester sandbox into general availability. Any developer can call it now through the Gemini API and AI Studio, and the Gemini app picked up a task mode that keeps running while you are away, wired to more than forty apps. The model that debuted locked to a handful of cyber defenders is a public endpoint today, with a 1M-token context window and a 1M-token output ceiling.

So the real question now is what a task costs. Here the sticker lies to you. Google is charging $2 per million input tokens and $10 per million output as an introductory rate, with no end date promised. The standard rate waiting underneath is $4 and $20. Cached input is discounted hard, which helps when you reuse context, but the figure that should trip a small alarm is this one: independent measurement puts Argon near 62,000 output tokens per task on average. At the standard output rate, one hands-off run is more than a dollar of output alone, before you count a word of what you fed it.

Task mode is where that math gets real. A mode built to keep working while you step away is, by design, a mode nobody is watching. Point it at a loosely scoped job and it can loop, reread, and regenerate on your card until it decides the work is done. The capability is real and the long-context retrieval scores are strong, so this is worth running. The discipline is to run it on a bounded task behind a hard spend cap, and to read the total task bill instead of the per-token line.

Ship it? TRY, on a leash. Give Argon one well-scoped job where a 1M-token window earns its keep, a document pile, a code migration, a research sweep, and set a spend cap before you let task mode off the porch. Measure the full cost of a finished task rather than the headline token price, then hold it against the model you run today. The promo makes it cheap to start. Your own numbers decide whether it earns a second week.

Sources: Gemini 4 Argon pricing breakdown; Google unveils Gemini 4 Argon.

Tour de Headlines

🏢 Workday turned Sana into the front door for its agents and shipped Agent Harness to GA. Sana is the new Workday Home, the one place its agents, tasks, and decisions surface, and Agent Harness is now generally available underneath it. The harness is the part worth copying: agents run in isolated environments with built-in memory, it governs third-party agents that reach data over MCP alongside the ones Workday built, and Sana coordinates all of them while following each organization's existing permissions and approval steps. Most readers here do not run Workday, so take the shape rather than the product. Every agent you leave running needs a data boundary, an approval checkpoint, and a record you can audit. Verdict: WATCH if you are a Workday shop, borrow those three lines if you are not.

🛠️ DXC put LangGraph into rooms where a wrong move gets audited. DXC is running its OASIS IT-operations platform on LangGraph for orchestration and licensing LangSmith for evaluation and observability, and says it is scaling toward 125 customer environments in regulated industries. Agents investigate incidents in parallel and pause for a person when an action needs approval. Count the 125 figure as DXC's own number until a customer says it out loud. The transferable piece is the operating rule they built around the framework: replayable evals for the incidents you hit most, a hard limit on what the agent may do on its own, and a named human in the escalation path before you widen its reach. Verdict: WATCH, and steal the eval-and-escalate habit even if you never touch OASIS.

🖥️ Rocket Software put agents on the mainframe. Rocket acquired Geniez AI, whose agents run straight on mainframes to watch systems, diagnose incidents, and carry out workflows from plain-English instructions, then folded the team into its EVA agentic platform. This is an acquisition, so there is no general-availability date to circle yet. It is also a reminder that agents are reaching the systems where an undo button does not exist. If you run one here, keep it in the passenger seat: let it monitor and diagnose, require a human sign-off for any change, and log every step. A nightly COBOL batch is a rough place to learn your agent was overconfident. Verdict: WATCH.

Tool of the Day

🧩 Skills Hub

Write an agent skill once and keep it in sync across every coding tool you use.

If you run more than one coding agent, you already have the problem this fixes. A skill you wrote for Claude Code drifts out of step with the copy in Cursor, in Codex, in OpenCode, until nobody remembers which one is current. Skills Hub (open source, MIT, shipped October 11) is a small desktop app that holds one source of truth and pushes it to all of them. You can read exactly what it writes and where before you trust it with your setup, which is the right bar for anything that edits agent behavior. If your skills live in four slightly different files right now, this is a twenty-minute cleanup that pays off every time you edit one.

Skills Hub on GitHub →

Worth a Click

• The honest Argon math. The $2/$10 promo, the $4/$20 bill underneath it, and what one finished task really runs. Read it before you turn task mode loose.
• Workday's announcement. Sana as the new Workday Home, in Workday's words. Read the guardrails and auditable-outcomes claims before you take the harness pitch at face value.
• SolidCraft. Parametric 3D CAD written in Rust with a built-in MCP server, so an agent can model a part through MCP instead of a mouse. Apache 2.0, and the most fun repo of the day.

One pattern runs under all four stories, and I will put a stake on it. The model layer is turning into a utility you call, cheap to start and easy to swap out. The value, and the risk, moved to the harness, not the model: the governed agent platform Workday shipped, the eval-and-escalate rule DXC built, the human sign-off a mainframe agent needs before it touches anything. A smarter model with no leash is just a faster way to make a mistake you cannot see.

So the cheapest thing you do this week is write three lines for every agent you leave running: what it can touch, what it can do without asking, and where its actions get logged. Argon's promo price makes it cheap to start one of these. Those three lines are what make it safe to walk away from. If I am wrong, it is because the governed suites turn out to be theater and a bare API call was always enough. I have not once seen an unwatched agent stay cheap or stay polite. I will write the three lines.

- Ron