The Product Desk

An ops lead on Business Premium who wants a vendor sheet watched does not shop for an "AI teammate." She types @ChatGPT in Slack. Per TheSequence's DevDay recap, Dots are rolling out now to Pro and Business Premium users in eligible markets, and the first Dot costs nothing extra. Enterprise gets a beta once a workspace admin switches it on. OpenAI also put @ChatGPT inside Slack and Teams and added ChatGPT Space and Pages for shared human–agent work. Computer use arrived in the Agents API, along with a Pro 500 tier carrying 25× the Plus usage allowance. Jesus Rodriguez's summary: "The chatbot is acquiring an office."

A Dot's persistence is what makes it a competitor. It runs on GPT-6 Astra on its own cloud computer and keeps working toward a goal between conversations. It watches for a change, chases a follow-up and moves data between apps. Many horizontal "AI teammate" products sell exactly that job list. It now ships inside a subscription the buyer already pays for, in the tool where the work happens.

Where the money says the defensible ground is

The financing rounds TheSequence reports fall on either side of the line Dots has drawn.

The two vertical companies are valued on revenue and named customers. The horizontal consumer agent is valued on expectation, and its job list now overlaps almost exactly with a bundled product. All three are financing-round valuations, not market prices, but they point the same way. A generalist agent can't quickly copy vertical depth, proprietary data or system-of-record status.

Re-run your cost model at two prices

OpenAI says GPT-6.1 Sol nearly matches Astra on agentic coding, computer use and professional work at about one-fifth of Astra's token price, by the company's own report. Teams tend to read the headline rate. An always-on agent spends its day re-reading the same context, so the line that sets the bill is cached input at $0.10 per million tokens, which is 20× cheaper than uncached. Cache-hit rate, the share of input served from cache, is the biggest cost lever you control.

Speed is now a tier of its own. Ultrafast mode generates up to 8× faster in Codex. It lands on Astra for Pro 500 and Enterprise first, and Sol's version is due within days. If some power users would pay to wait less, check whether the pricing table has a row for them.

The new price may not hold. Google matched it with Gemini 4 Argon the same week, and The ML Engineer notes that Argon's rate is introductory and set to double to $4/$20. Model any feature that only pencils out at the new rate at Sol's list price and again at twice that, the same jump Argon already has scheduled.

The move

Before the next roadmap review, list the top ten user jobs and ask of each whether a Dot could do it inside Slack or Teams today. We have argued here before that the tool already open on the buyer's screen usually wins the tie. Every "yes" needs one of two answers. Either you hold something a generalist can't reach, such as proprietary data, vertical depth or the system of record, or you become a connector Dots call into instead of a feature they replace.

When an agent misfires, the cheapest fix teams reach for is a second prompt that asks it to check its own work. NVIDIA and KAIST's Mid-Harness study tested that fix directly, and it mostly failed. The researchers also distilled the frontier verifier's judgment into the small model, which means training it to imitate the stronger checker. Even then, self-checking reached only 57.14%, well short of what the separate checker achieved. A stronger, independent checker is what moves the number.

The setup is simple to copy. A 9-billion-parameter agent proposes eight candidate actions for each step. A stronger model, GPT-5.6 Sol, picks one before anything executes. The generator and the surrounding harness stay untouched. Adding best-of-three full attempts on top produced 66.33%, no better than the checker alone. TheSequence argues that at frontier prices of $2/$10 per million tokens, a design where a small model acts and a frontier model verifies is now affordable for most mid-value actions.

Your eval wins may be the part that overfits

RRSI comes from Google Cloud AI Research with Stanford, UNC-Chapel Hill and WashU in St. Louis. It studies what happens when an LLM keeps editing the prompts, tools and control flow around a frozen model, here Claude Opus 4.8, against a single evaluation set. The gains on that set shrink or vanish on new tasks. As The ML Engineer reports, RRSI's regularized version posted the smallest gain on the set it tuned against. It also posted the best out-of-distribution average: 43.6, against 39.7 for the base harness. TheSequence adds that all six held-out benchmarks improved, by +3.4 on average. Both outlets report roughly a third fewer tokens per trial, so the cost side checks out too.

Three mechanisms did the work:

• An edit budget that shrinks from bundled changes to single changes over the run.
• A critic that rejects any edit encoding task names or benchmark answers.
• A pruner that flags components that have stopped earning their keep.

For a PM, the translation is uncomfortable. If your team iterates prompts against one golden set and reports a win rate, that win rate is the number most likely to mislead you.

The production version of the same idea

Anthropic's engineers applied the pattern to performance work. Per Arc Notes Weekly, agents in a two-week sprint cut the time until a user can type on a fresh claude.ai page from 3.1 seconds to 0.55 seconds at p75, the experience of the 75th-percentile user. The agents worked against reliable benchmarks. Each gain was checked against real-user latency, and each win became a CI check, an automated test that blocks regressions. Humans set priorities and rejected complexity that wasn't worth the milliseconds. Anthropic's headline says "3x"; the raw p75 figures are the safer ones to cite.

All three results share a pattern. The gain came from a measurement the agent didn't control: an outside verifier, a held-out test set, or real-user latency. The research numbers come from specific benchmark suites, so measure the lift on your own evals before you promise it in a PRD. The product consequence is that the success metric becomes the spec. If the goal is vague, agents will ship a vague product, only faster.

The most telling detail in Uber's write-up is how the tools get made. A workflow called AutoCrawler continuously scans Uber's API definitions and generates MCP servers from them, and an LLM writes the tool descriptions. That is how a company reaches thousands of tools quickly. It also means the tool count outruns anyone's ability to review it, and that is why Uber's first control is the default state, per The ML Engineer's issue #407.

Uber's architecture, read as product requirements

The cost row needs explaining. Context bloat means the agent's working memory fills with tool lists and raw outputs it doesn't need, and you pay for every one of those tokens. Uber's fixes address it three ways. Tools are shown only as the agent needs them, the agent can request specific fields, and large outputs go to files instead of into context.

"On whose behalf" has no standard answer yet

The same author's two-part series uses Alice and Bob, two users of one agent platform who delegate different actions to their own agents. It traces their requests across agent graphs using KAOS and the new open-source Zalando Agentic Identity Broker. Most PRDs skip two of its cases. The first is how the agent's own calls to its tools and models get authorized. The second is what an autonomous agent with no user behind it is allowed to reach. The author expects agent identity to stay unsettled for years, so your team will make some of these calls itself.

Even the model's builders keep a human gate

Anthropic's Alex Palcuie describes using Claude for incident response as having moved from "no" to "maybe," per Sylvain Kalache's talk summary highlighted in SRE Weekly. The vendor's own engineers say "maybe" on production fixes, so any roadmap that promises autonomous remediation is ahead of the evidence. The same issue points to Netflix's Stratum right-sizing system as the model to follow. Stratum automates, but it ships with guardrails, safeguards and an automated fallback.

Lorin Hochstein's incident analysis adds a warning for agents that run your runbooks. Failures come from correctly following a wrong procedure or incorrectly following a right one. An agent can execute a wrong runbook perfectly and fast.

The market is moving the same way. NVIDIA launched an Open Agent Safety Platform whose Sentry component quarantines agent breakouts in milliseconds, per TheSequence. Anthropic, Microsoft, JPMorganChase and 100+ other organizations back it. The Information reports that agents "going rogue" are heading toward novel legal battles, and a federal AI task force led by Jay Clayton will report within 120 days. Buyers will soon expect agent permissions the way they expect SSO.

The practical translation is three autonomy tiers for every agent action: suggest, act with human approval, and act with automated revert. Allow the third tier only where a tested fallback exists.