AI Workflow Weekly Vol. 1

The big picture

Codex prototyped an agent that keeps working until you stop it. Cisco moved forward with deployment to 90,000 employees. AccuKnox productized agent permission control. Meanwhile, Zhipu and Alibaba released low-cost models, and a 15-person Indian startup raised $21M to push AI past building into revenue growth. The thread connecting all five: demand for the infrastructure around AI, including stopping, managing, auditing, cost control, and monetization, is growing faster than AI capability itself.

This week's five stories worth watching

• 01

  Codex "Persistent mode" code surfaces

  Aug. 27, WIRED; code merged to GitHub on Aug. 26

  WIRED found Persistent mode in Codex's open-source CLI codebase. The setting lets the agent keep working until a user explicitly puts it to sleep, create its own follow-up tasks, carry them across sessions, and message the user unprompted. OpenAI confirmed it is experimenting but has no near-term launch plans. The API currently receives only a disabled value.

  ANOS take: Persistent AI is useful, but it creates risk if it runs without control. OpenAI itself has acknowledged cases where runaway agents tried to break out of their sandboxes. Stop conditions, cost ceilings, and audit logs must come before real deployment.

• 02

  Cisco finishes deploying "MyAgent" to 90,000 employees

  Announced Jul. 1 by Fortune; deployment reported in late Aug. by WSJ

  Cisco's personal AI agent MyAgent is now live across its workforce of about 90,000 people. It routes each task to the most cost-efficient model rather than defaulting to a frontier model. CFO Mark Patterson says AI already generates 80 to 90% of the first draft of the MD&A section in Cisco's public filings. AI-related orders were $2B in FY2025, with guidance raised to $9B for FY2026. Cisco also laid off about 4,000 people in May, so AI deployment and workforce reduction are running in parallel.

  ANOS take: The real story is not simply that 90,000 people got an AI. It is the routing architecture: different models for different tasks, with cost as a first-class variable. That design pattern matters more than the brand name on the agent.

• 03

  AccuKnox launches AgentZ

  Aug. 27, GLOBE NEWSWIRE

  AgentZ is a model-agnostic platform that bundles agent execution, tool connections, role-based access, runtime credential injection, workflow orchestration, audit logs, and sandboxing into one stack. It deploys as SaaS, on-prem, or air-gapped. The company comes from the SRI International zero-trust security ecosystem.

  ANOS take: This may be the most underappreciated story of the week. The industry is still debating what agents can do. Enterprise buyers care about what agents must not do. This category is likely to expand over the next 12 months.

• 04

  GLM-5.3-Flash and Qwen3.8-Flash-Next released on the same day

  Aug. 26, Zhipu AI (Z.ai) / Alibaba Qwen

  GLM-5.3-Flash: a MoE model using 18B active parameters out of 320B total, with open weights under MIT. It was previously stealth-tested as "Ox Alpha." Pricing is $0.15/M input, or $0.075 during promotion. Qwen3.8-Flash-Next: activates 6B out of 125B parameters and previews the next Qwen4 architecture. The production API, Qwen3.8-Flash, runs at $0.16/M input and $0.47/M output. Low-cost model options are expanding quickly.

  ANOS take: If your revenue depends on API markup, this is a threat. If your revenue depends on the product you build on top of the model, this is good news. Model commoditization is accelerating; differentiation is shifting from which model you use to what problem you solve with it.

• 05

  Runable raises $21M Series A

  Aug. 26, TechCrunch / Business Wire

  Bengaluru-based, 15-person startup Runable closed a $21M Series A co-led by Susquehanna VC and Nexus VP at a $65M post-money valuation. It went from $0 to $2M ARR within three weeks of launching payments in March and has about 1.7M registered users. The product goes beyond building sites and apps: it handles SEO, social ads, ChatGPT Ads, cold outreach, and AI search optimization, covering the full path from creation to customer acquisition.

  ANOS take: Fifteen people, 1.7M users, and $2M ARR in three weeks show the leverage AI-native teams can now have. The key idea is that the agent that builds your product can also help market it.

Five questions to ask your team on Monday

To turn this week's news into action, not just reading material, here are five questions to ask on Monday morning. If any answer is No, that is your current bottleneck.

• 01

  Do you have a kill switch?

  From Codex. If you run agents for hours, you need a cost ceiling, a time limit, and a hard stop. Letting it run is not a strategy.

• 02

  Who gets access to what?

  From Cisco. Giving everyone AI multiplies your data exit points. Design the access controls before the rollout, not after.

• 03

  Can you trace what agents did?

  From AgentZ. Which tools did the agent call, what data did it touch, what did it send externally? An agent you cannot explain will fail audits.

• 04

  Are you routing by task?

  From GLM/Qwen. Use frontier models for hard decisions and low-cost models for repetitive processing. Routing by task can change costs dramatically.

• 05

  Are you measuring outcomes?

  From Runable. Not tokens generated, not pages built. Inquiries, conversion rates, and revenue impact decide AI ROI.

If even one answer is No, fix that before adding another AI tool.

Products you could build from this week's news

Where the big players advanced, gaps opened for smaller AI companies. Five concrete ones from this week.

• 01

  Agent operations dashboard

  As persistent agents become real, teams need a single view showing each running agent's status, token burn, cost, and a stop button. Dedicated SaaS in this area is still limited.

• 02

  SMB AI deployment kit

  Cisco can build its own routing infrastructure. Companies under 100 people usually cannot. Package model routing, access controls, and usage logs together.

• 03

  Lightweight agent audit log

  AgentZ is enterprise-grade. A simpler log-and-search service for what agents did would fit mid-market compliance needs such as ISMS and SOC 2.

• 04

  Multi-model cost optimizer

  As low-cost models such as GLM-5.3-Flash and Qwen3.8-Flash expand, automatic routing and cost visualization by task type can become a product on its own.

• 05

  Build-to-revenue pipeline

  Runable's direction: one agent builds the site, then supports SEO, ads, and outreach. The Japanese market still has room here.