This Week’s Cheaper AI Models Push Agent Skills Beyond Coders

This week’s AI news answers one question directly: agent skills are worth more to legal, recruiting and sales teams right now than to engineers. Google halved the price of its coding and agent model, Claude Code began running more tool calls without asking first, and OpenAI’s own data shows non-engineering teams adopting agents fastest.

What is Gemini 3.7 Flash for?

Google released Gemini 3.7 Flash on August 13, 2026, built for coding and for AI agents that carry out multi-step tasks such as writing code, running it, and fixing what breaks, according to Google’s announcement. Google says developers get more reliable multi-step coding and agent task completion as a result, per the same source. DeepMind’s own model card adds that the new Flash scores higher on coding and terminal-agent benchmarks than the model it replaces. A terminal agent is one that can open a command line, run a program, read the error, and try again on its own, closer to how a junior engineer debugs than how a chatbot answers a question. That distinction is why this release fits agent workflows, not just chat replies.

Why is the price cut the bigger story?

The number worth remembering here is the price, not the benchmark score. Gemini 3.7 Flash costs half of what the previous Flash model cost, and Google confirms that price stays through the end of 2026. DeepMind’s model card repeats the same figure, half the price of the prior Flash model, with stronger coding and agent scores attached. I wrote about this pattern when earlier price cuts gave workers leverage rather than just savings for a finance team. The same logic applies here. When the cost of running an agent for a full afternoon drops by half, a team does not need a bigger budget approval to try it. They need a task worth automating and someone willing to check the output.

What changed in Claude Code this week?

Anthropic switched Claude Code to auto mode by default for most sessions on Pro, Max and Team plans starting August 14, 2026, according to its auto mode documentation. Under the old default, Claude Code stopped and asked before most tool calls, actions like editing a file or running a command. Under auto mode, those calls run through a safety classifier that decides whether to proceed without asking, based on Anthropic’s own configuration guide. For anyone coding with Claude daily, that changes the rhythm of the session: fewer stops, more autonomous action, and a tool making more calls about what counts as safe before a human sees them.

Is auto mode safe to trust by default?

Auto mode moves judgment earlier in the process, into the safety classifier Anthropic built and describes in its auto mode documentation. That documentation confirms auto mode applies to most Pro, Max and Team sessions, but it does not publish a rate for how often the classifier misjudges a risky call in your specific codebase. If your manager already doubts AI-written code, an agent that now decides more for itself makes that conversation harder to have well. This is what I train people to say when a skeptical manager pushes back: explain AI-assisted work to a boss who doesn’t trust it.

Where is agent adoption growing fastest?

OpenAI’s own usage data, reported by VKTR on August 13, 2026, shows its top enterprise users now consuming eight times more than a typical firm. The same report shows legal, recruiting and sales teams adopting AI agents faster than engineering teams do. That is worth sitting with if you assumed agent skills were an engineering problem. The people closing deals, screening resumes and reviewing contracts are running more agent tasks according to VKTR’s account of the data, not fewer. That is a specific place to point your own learning time if you work outside a codebase.

What does this data not prove?

This is OpenAI’s own usage data about its own paying customers, not an independent audit of who uses agents best, per VKTR’s report. It measures token consumption and account activity, not whether the work an agent produced was accurate or checked by a person before it went out the door. It also excludes every company running agents on Gemini, Claude or an open model instead of OpenAI’s, so “top enterprise AI users” describes OpenAI’s customer base, not the whole market. I made a related point when agents got cheaper this week, not safer: usage volume tells you adoption, not quality. Treat this data as a hint about where to look, not proof of where to spend your training budget.

Four roles gaining agent ground, ranked by adoption speed

Based on OpenAI’s data as reported by VKTR on August 13, 2026, here is how four groups compare on agent adoption this week.

1. Legal teams, named among the fastest-adopting groups in the OpenAI data VKTR reported.
2. Recruiting teams, using agents for screening and coordination work at a pace close to legal’s.
3. Sales teams, the third group VKTR’s report names as surging ahead of engineering.
4. Engineering teams, still adopting agents but at a slower pace than the three groups above, even as Gemini 3.7 Flash and Claude Code’s auto mode are built specifically for them.

What should you do this week?

Pick one recurring task in your job, something you do the same way every week, and ask whether an agent could produce a first draft of it while you check the result. If you work in legal, recruiting or sales, this week’s data says you are already ahead of the curve, per VKTR’s report. If you code, the cost of testing Gemini 3.7 Flash against your current setup just dropped by half, and Claude Code will now act on more of your instructions without waiting for a yes, per Anthropic’s documentation. Either way, the skill worth building this month is judging an agent’s output quickly, not writing the perfect prompt.

You are a workflow consultant helping me find one task in my job that I could hand to an AI agent this month. Here is my role: [describe your job title and 3-5 tasks you do every week]. I want a task that is repetitive, has a clear right answer I can check quickly, and does not involve sharing confidential client or employee data. Ask me clarifying questions about my week before you answer. Then give me one specific task to try, the tool or approach you'd use, and exactly how I would check the agent's output before trusting it.

You are a senior engineer reviewing my Claude Code setup before I turn on auto mode by default. Here is my project: [describe language, whether it touches production data, and whether it has automated tests]. List the specific actions in a coding session I should still want to approve manually, like deleting files, pushing to a remote branch, or changing environment variables, even under auto mode. Then tell me what settings or guardrails in Claude Code's auto mode configuration I should check before trusting it on this project. Ask me questions about my setup first if you need more detail.

You are helping me decide whether to switch a coding or agent task from my current model to Gemini 3.7 Flash. Here is the task: [describe what the agent does, how often it runs, and roughly how much it currently costs per month]. Compare likely cost at half the price of the prior Gemini Flash model against the risk of lower output quality on this specific task. Tell me what to test on a small sample before switching everything over, and what result would tell me the switch was a mistake. Ask clarifying questions about my task before answering.

Act as a blunt hiring manager who has read ten thousand resumes, not a career coach. I will paste my full resume and the job description I want. Rewrite the whole resume for that role, section by section, in this order: summary, experience, skills, education. Rules: every experience line leads with impact, not duty. Use bracketed placeholders like [8 percent] for any number I did not give you, and list at the end every placeholder I need to replace with a real figure. Keep it to one page of text. Plain formatting only, no tables or columns, so screening software can parse it. After the rewrite, tell me the three weakest claims that need evidence before I send this anywhere. My resume: [paste resume]. The role: [paste job description].