Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

• Charlie Labs, a startup building an AI engineer, shut down — learn why below

• Who gets paid when AI does the shopping?

• OpenAI DevDay was full of new announcements — check the most important ones below

Today’s newsletter is sponsored by Thatch, the employee health benefits platform helping companies offer flexible health budgets.

The 2027 Open Enrollment Playbook: timelines, checklists, and templates ^AD

Inside the 2027 Open Enrollment Playbook: a week-by-week timeline, a communication plan that cuts down support tickets, employer and employee checklists, and how to weigh your group plan against an employee health budget with fixed, predictable costs.

Renewal shock, confused employees, endless admin: open enrollment doesn't have to feel like this. Get the free step-by-step playbook benefits leaders are using to plan for 2027.

→ Download now

This Week In Startups

🔗 Resources

How to create your own personal AI benchmark

Apps, Agents, and Aggregation

Does concentration risk matter in AI?

Who gets paid when AI does the shopping?

📰 News

Anthropic released Claude Sonnet 5.5

AMD is acquiring World Labs

Nvidia launches new platform for reining in rogue AI agents

Shopify debuts Canvas, a way to build online stores by chatting with AI

💸 Fundraising

EliseAI, a startup that automates leasing and patient scheduling for healthcare businesses, raised $350M.

Flow Engineering, a startup that uses AI agents to help engineers design and test hardware, raised $50M.

Reverion, a startup that builds reversible fuel-cell power plants, raised $175M.

Armadin, a startup that uses AI agents to find exploitable weaknesses in company systems, raised $255.5M.

Fail(St)ory

One More AI Engineer

Charlie Labs built an AI engineer that handled coding tasks and recurring maintenance inside GitHub, Linear, and Slack. It announced its closure on September 21, with the service ending October 5.

Charlie tried to earn its keep by cleaning up after everyone else’s coding agents. Now another AI engineer is heading for the exit, after trying to sell developers an extra teammate while Claude Code and Codex were already at work.

What Was Charlie Labs:

Charlie wanted teams to delegate work through the tools where they already discussed it. An engineer could assign a Linear issue, work through the plan, and get a pull request back. Follow-up comments became instructions for the next change.

The agent ran in the cloud, with its own development environment and access to connected repositories. Work could continue after the developer closed their laptop, and the discussion stayed in GitHub, Linear, or Slack for teammates to inspect and continue.

Charlie could connect an error, an issue, and the relevant code across repositories. The team was buying someone to carry a task through its existing workflow, including the checks and review, with less supervision along the way.

In April of 2026, Charlie’s founder admitted that coding agents had advanced faster than expected. Therefore, the team decided to shift its emphasis toward maintaining the work those agents produced: reviewing changes, updating documentation, investigating alerts, and keeping pull requests ready to merge.

It packaged those jobs as Daemons. Teams committed small files to their repositories describing each role and when it should run. A documentation daemon could check for drift on a schedule; another could investigate incoming bugs. Teammates could review and change those instructions alongside the code.

Charlie supplied the hosting, event handling, and execution behind each role. Engineers still reviewed outputs and decided what the agent could access, but they could delegate recurring chores without starting a fresh conversation every time. Coforma reported 1,198 merged pull requests over eight months, using Charlie across dozens of repositories.

The company sold that capacity through workspace plans with unlimited members and paid overages. More delegated work could mean a larger bill. As existing coding tools added background workflows, Charlie needed its shared context and maintenance roles to save enough additional effort to justify that bill.

The Numbers:

• 🚀 General availability: August 2025

• 💵 Paid plans: $50–$1,000/month, plus overages

• 📉 Repository-selection cost: 90.4% lower per call, company-reported

• 🧾 Usage-limit enforcement: July 14, 2026

• 📅 Service ends: October 5, 2026

Reasons for Failure: 

• The original workflow was a harder sell: Charlie acknowledged it had misjudged adoption. Developers mostly worked one-to-one with agents and kept using local tools. Getting a team onto Charlie meant changing how it assigned work, granted access, and reviewed results. That likely raised the effort required to win customers already comfortable with their own coding assistant.

• Existing agents kept expanding into its jobs: Anthropic introduced parallel, long-running workflows in May, and its cloud routines cover recurring reviews and documentation upkeep. Charlie’s maintenance focus put it alongside features of a broader coding product. Its repository-owned roles needed to deliver enough extra value to keep a separate subscription attractive.

• Recurring chores had to earn their bill: Charlie paid for model calls and cloud execution, including reviews that found nothing. Many customers had used it without usage charges or enforced limits before July. Enforcing those limits made the cost of delegation more visible, leaving customers to judge whether each recurring job saved enough time to keep running it.

Why It Matters: 

• Expanding tools like Claude Code and Codex make standalone AI engineers harder to justify.

• Heavy agent usage can overstate demand when customers haven’t faced the full bill.

Trend

DevDay Recap

OpenAI made more than 20 announcements at DevDay on Tuesday, including a persistent personal agent, a cheaper model, and new ways to build software inside ChatGPT.

Here’s a recap of the highlights, including a few that connect to trends I’ve covered over the past month.

Why it Matters

• Persistent assistants give users ongoing work to delegate, while decision models let developers assign small, repeated judgments to a specialized tool

• For software founders, OpenAI is also offering richer interfaces inside ChatGPT and a purchasing route tied to customers' existing spending commitments.

Dots

Dots are personal agents powered by Astra, with their own cloud computer, browser, memory, and ongoing goals. You can give a dot a responsibility and return to it over time. It keeps context and works between conversations, with ChatGPT, Slack, and Teams providing ways to steer it.

I think this is the most interesting announcement because it's another entry in the personal-assistant category we covered two weeks ago, alongside Meta's Muse and Instinct AI. The common idea is a continuing relationship with an assistant that remembers your priorities and follows up on them.

OpenAI’s examples include a dot monitoring customer feedback, building and testing fixes, and preparing code changes for review. For a product launch, it can update the plan and rewrite launch materials as the scope changes. The idea is to hand it an ongoing project and have it keep the work moving as new information comes in.

OpenAI also shared an early tester’s experience: his dot noticed he’d forgotten to invoice a publication and prepared the invoice for him. You can open your dot’s computer to see what it’s doing, add another task while it’s working, or call it to talk through a decision.

Decisions API

Decisions API uses Luna to answer questions with a predefined set of possible answers. It accepts text or images and can classify content, route requests, or choose an agent's next action.

For example, a support workflow could use a bounded decision to select the queue a request belongs in. That's a small judgment repeated many times inside a larger process.

This is the space we discussed last week with Jev. OpenAI's entry makes the separation of tasks Jev is built for: a dedicated model handles the decision, and other tools handle the work around it.

Plugins and Marketplace

OpenAI launched ChatGPT plugins in 2023 and the GPT Store in 2024. The original plugins were retired that year. Introducing this part of the keynote, Altman acknowledged mixed results from earlier attempts.

OpenAI is ready to try again. The new plugin extensions let developers build sidebar apps, interactive panels, and file viewers or editors. Users can inspect and change a product's work through its interface inside ChatGPT.

Alongside that, OpenAI Marketplace lets eligible enterprise buyers allocate part of an existing OpenAI commitment to approved partner software. The launch cohort includes Figma, Harvey, and Salesforce.

Space and Models

ChatGPT Space brings shared files and Pages into one place for people and agents. Pages are editable documents with comments and interactive content, so colleagues can work on an output together and keep it available for the next task. Collaborative slides and spreadsheets are still forthcoming.

Finally, there was also a new model: GPT-6.1 Sol. OpenAI reports near-Astra coding performance at one-fifth the standard short-context input and output token prices. You can now also use Astra’s new Ultrafast tier which offers faster token generation in Codex, with higher usage charges.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico