self.md radar — 2026-05-12
Agent adoption got dragged into management plumbing today.
GitLab tied restructuring to the agentic era. Shopify and Notion treated the shared work record as the agent interface. Anthropic and AWS gave enterprises native Claude features with a boundary footnote big enough to matter.
1. GitLab made the org chart part of the agent plan
sources:
what happened: GitLab published its Act 2 letter on May 11 and said the agentic era is the largest opportunity in the company’s history. The same letter announced a workforce reduction and tied four operational changes to it: reducing the number of countries with small teams by up to 30%, rewiring reviews, approvals, and handoffs with AI agents, and reorganizing R&D into roughly 60 smaller teams with end-to-end ownership. It is also retiring CREDIT, the values framework behind Collaboration, Results for Customers, Efficiency, Diversity, Inclusion & Belonging, Iteration, and Transparency. Simon Willison noticed the messy denominator: GitLab says it operates in nearly 60 countries, while the public employee handbook lists 18.
why this matters: This is not an IDE feature announcement. GitLab is making a claim about company shape: if agents can remove handoffs, the map of teams, countries, approvals, and management layers changes too. The uncomfortable part is that the same sentence can describe real process cleanup or a layoff deck with better nouns.
2. agents got pushed into the public work record
sources:
what happened:
Tobias Lütke described Shopify’s internal coding agent River as a Slack-native worker that refuses direct messages and asks people to create public channels instead. Lütke says he works with it in #tobi_river; the point is Lehrwerkstatt-style learning by watching the shop floor, not private prompting. On the same day, Lenny’s How I AI published Ryan Nystrom’s Notion workflow: automate standups, ship PRs from a single comment, and use spec-first development so agents do the coding while the human holds the shape of the work. Nystrom has been a core builder of Notion AI and the Custom Agents feature launched in February 2026.
why this matters: The useful interface is the work record. A public channel, spec, or PR comment gives the next human something to inspect; a DM gives them folklore and a transcript nobody wants to read. Agents are easier to trust when their work leaves a trail other people can learn from.
3. Claude on AWS split access from custody
sources:
what happened: AWS announced general availability for Claude Platform on AWS on May 11, calling itself the first cloud provider to offer the native Claude Platform experience. Customers get AWS IAM credentials, access controls, consolidated billing, and CloudTrail audit logging, plus Anthropic’s platform surface: Claude Managed Agents beta, advisor strategy beta, web search, web fetch, code execution, Files API beta, Skills beta, MCP connector beta, prompt caching, citations, batch processing, and the Claude Console. The catch is printed plainly in the AWS post: Claude Platform on AWS is operated by Anthropic, and customer data is processed outside the AWS security boundary. Anthropic says Bedrock remains the path where AWS is the data processor, while this route gets same-day access to native Claude features and can retire against AWS commitments.
why this matters: This is the enterprise compromise: keep procurement and audit habits inside AWS, move the product surface to Anthropic. For agents with files, code execution, skills, and MCP, that boundary is not legal wallpaper. It decides where the run can be inspected, billed, and blamed.
supporting links
- Artificial Analysis Coding Agent Index methodology — the new public index combines SWE-Bench-Pro-Hard-AA, Terminal-Bench v2, and SWE-Atlas-QnA, with pass@1 plus cost, token, and execution-time accounting.
- Elastic on PHANTOMPULSE in Obsidian — a shared Obsidian vault, Shell Commands, and Hider became the delivery chain for a cross-platform RAT; the PKM layer is executable now.
- manifest
— a model router for agents with 300+ models, 16 providers, a shared
/autoendpoint, cost tracking, and fallback rules. - Puter — the self-hostable internet OS keeps showing up in the personal-OS lane; GitHub’s API showed 41,100 stars and an AGPL-3.0 license this morning.
left on the table
- AWS’s OpenAI-on-Bedrock move was the recent cousin. The Claude Platform item stayed because this is a GA service with a different custody line: Anthropic operates it, outside the AWS security boundary.
- Artificial Analysis did not become the main signal because agent benchmarks and accounting already carried May 9 and May 10.
- manifest stayed supporting-only because routing was already the May 4 lead; this repo is useful, but not a new thesis by itself.
- the NYT AI quote correction was ugly provenance evidence, but today’s stronger pattern was inside the systems that assign work, spend money, and keep logs.
Related self.md guides
- Best Claude Code plugins — choose the Claude Code extensions worth installing, and the ones to skip
- Best MCP servers — connect files, browsers, memory, search, and workflow tools without turning the stack into soup
- Sandboxing and security — give agents useful power without giving them the whole house