Plenty of AI sessions cover what agents can build. This one covers the other side: what happens when an agent takes the wrong action, and the controls that prevent it.

In this session, we’ll cover:

• The common ways agents fail, and what each one costs

• Where to put a human approval step, and where it just adds friction

• Limiting what an agent can access and reviewing what it did

• How to decide which work stays with a person

If you have held off on agents because you were not sure you could keep them in check, this session is built for you.

Tuesday, August 25, 2026 · 12:30 PM ET / 11:30 AM CT / 9:30 AM PT · Free live webinar with Josh Sullivan, COO at Kiingo AI.
Register for the August 25 webinar →

• OpenAI made GPT-5.6 Luna the default model for free ChatGPT accounts on August 6, removed the cap on text conversations, and added a Think button that spends longer on harder questions. Four days later it released GPT-5.6-Cyber and restricted it to vetted defenders through a two-tier program called Daybreak, saying the model completes 95% of tasks on its internal cybersecurity evaluation against 1.5% for standard GPT-5.6 Sol. Uploads, image generation, and tools stay metered on the free tier. (OpenAI, TechCrunch, Axios)

• Starting August 14, Anthropic will run Claude Code in auto mode by default on Pro, Max, and Team plans, so it proceeds without asking unless a step is judged irreversible, destructive, or aimed outside your environment. Anthropic cited its own study of 1,053 paid testers, which found people approve 97% of the permission prompts they see. Enterprise and API customers follow within a month. (Anthropic, TechCrunch, InfoWorld)

• Six companies published Agent Plugins on August 6, an open standard that packages agent skills and server configurations into one folder any compatible client can read. OpenAI, AWS, GitHub, VS Code, Cursor, and Vercel are behind it. Google has since joined as a maintainer. The pitch is building a plugin once and running it in Codex, ChatGPT, Cursor, and Copilot without rewriting the setup for each one. (Agent Plugins, Vercel, Google)

• Meta shipped two developer products in six days: Muse Code, a terminal coding agent, on August 5, and Muse Glimmer, a 30-billion-parameter model under an Apache 2.0 license, on August 10. Glimmer runs on a single 24GB consumer graphics card and the license permits commercial use. Muse Code is billed at $1.25 per million input tokens and $4.25 per million output. (CNBC, TechCrunch, VentureBeat)

• Google said the Gemini app passed 1 billion monthly active users on August 11, up from 950 million in July. The company says 63% of Gemini users use voice and the app now generates more than 150 million images daily. (Google, TechCrunch)

• Akamai’s Enterprise AI Usage Risk Report, released August 5, found that 47% of enterprise AI conversations run through personal accounts rather than company-managed ones, and more than 14% go through personal freemium subscriptions. The most active 5% of users generate the majority of prompts, and almost 75% of AI browser extensions request high or critical permissions. (Akamai, Akamai SOTI report)

• Lovable raised $400 million at a $13.3 billion valuation on August 12, doubling its valuation since December. The company says users have created more than 60 million projects, with applications built on its platform receiving over 900 million visits each month. (Lovable, TechCrunch, Axios)

• Bloomberg reported on August 6 that swings in AI power draw are wearing out data center batteries, generators, and cooling equipment ahead of schedule, with fluctuations exceeding design capacity by up to 50%. Operators are treating equipment budgeted as long-lived capital assets as consumables. (Bloomberg, Fortune)

• Anthropic said on August 10 that an unreleased version of Claude raised the proven lower bound for zeros of the Riemann zeta function satisfying the hypothesis from 41.6% to 67.2%, the largest single step on that constant to date. The run used about 60 coordinated subagents, 31 million output tokens, 2,400 shell commands, and a review of 54 arXiv papers across two sessions. Anthropic says it does not expect these techniques to prove the hypothesis itself. (Anthropic)

• Two-month-old River AI raised $1.1 billion across its seed and Series A rounds, led by General Catalyst and AMP PBC, with Nvidia and AMD participating. River is building tools that let companies train and own customized open models without maintaining a dedicated AI infrastructure team. (River AI, TechCrunch, Axios)

• BCG asked 152 CEOs at $500 million-plus companies and found 88% see cost or revenue benefits from AI in targeted areas. The companies turning those into enterprise-wide gains are 7x more likely to redesign workflows end to end and 2.4x more likely to put their best talent on AI work. (BCG, July 22, 2026)

• EY surveyed 534 senior business leaders across ten industries and found 87% have deployed or are piloting programs to build their own internal software with AI, and 94% say it beats traditional development on speed. 93% say a governance framework has to come with it. (EY AI Pulse Survey, July 28, 2026)

You give an AI assistant a clean brief, a few examples, and a careful list of dos and don’ts. Then the draft comes back technically fine, and you rewrite it anyway. The assistant only ever sees what you asked for, not the judgment you apply when the work lands. Your real standards show up somewhere else: in the drafts you quietly rewrite. Show it the work you rejected and it can work out the standard you are actually applying.

1. Pull two or three drafts you turned down. The close-but-not-good-enough ones: the AI draft you rewrote before sending, the agency draft that sounded polished but not like you, the summary that got every fact right but missed the point. A paragraph from each is plenty.

2. Add one line on why each failed. Too long. Buried the ask. Sounded like a press release. Written for the wrong reader.

3. Paste them in with this:

“Here are drafts I rejected, each with a note on why. Infer the standards I am actually applying and list them back to me as rules. Then rewrite [what you are working on] against those rules, and flag any rule you were unsure how to apply.”

4. Correct the rule list. Read it before you read the rewrite. It will get a few rules wrong, and fixing those lines is the point of the exercise.

5. Save it where the assistant will keep reading it. Put the corrected list in a document called What Good Looks Like Here. In Claude, open your Project and add that document to project knowledge. In ChatGPT, open your Project and upload it to the project’s files. In Gemini, open your Gem and paste the list into the instructions field.

Then add one line to that project’s own instructions: “Before delivering anything, review every document in this project and apply the standards in them.”

From then on, anything you start in that space comes back already matching your standard. The same document is what you hand a new person on day one.

I’m a multi-instrumentalist. Everything I play is loop-based, with effects placed deliberately at different points in the chain, and for the past few months I’ve been building a completely improvised solo project to perform live. Nothing is written in advance. That’s the whole premise.

An improvised set has more moving parts than a rehearsed one, because every decision has to hold up in real time. Signal routing and what feeds what. Power and levels. Building and dropping loops while switching instruments mid-piece. What I do when something falls out in front of an audience. I had been carrying all of it around in my head at once.

Working through it with AI, one decision at a time, is what got it organized. It helped me map the full chain, think through the points where it could fall apart, and turn a pile of loose intentions into a setup and an approach I could rehearse against. That structure is what moved this from sessions in my practice space to something stage-ready.

The improvising stayed exactly as it was. It’s still unwritten, still the part I care about most. AI has been instrumental in everything around it, pun fully intended, hehe.

Your AI tools are taking more routine work off your plate by default. A quick review of those settings helps you decide where that extra autonomy will save your team the most time.

Kiingo makes your company AI native: a sequenced roadmap of what to automate next, practical training for the people doing the work, and the approvals and limits that make each step safe to take.

Let’s talk about your roadmap →