AI work is standardizing the workflow and splitting the org chart
Databricks and Paradev are turning AI workflows into reusable assets. Engineering teams, meanwhile, are reorganizing in opposite directions.
- Skills in Genie One are written in plain language, can be chained together, and can be published through Unity Gateway for any agent supporting MCP. m:…8106
- A pull request in a stack now merges with the layers below it, and whole stacks can join the base branch's merge queue. m:…5089
- 40.8% of surveyed engineering leaders reorganized around products or outcomes, while 29.6% changed nothing at all. m:…7679
- Almost 90% of directors reported some form of agentic coding in their interview loops. m:…7679
Skills turn workflows into reusable assets
Skills in Genie One ensure a consistent format and output for tasks, and they are written in plain language and the open Skills format. A skill can be built in Genie One through a single conversation, and skills can be chained together to perform complex tasks. m:…8106
Skills can be published through Unity Gateway for use by any agent supporting MCP, and Unity Skills (Beta) provides a central place to publish and govern skills. Genie One learns business definitions and how teams use data through Genie Ontology, while Genie Agent acts as a domain expert for a specific organizational area, with the skill providing the task-specific workflow. m:…8106
The idea is that the workflow becomes the asset, not the model call. A team that writes down how a report gets built can reuse the method every time, and govern it in one place, instead of re-explaining the job to the agent on every run. m:…8106
Stacked PRs change how code gets reviewed
Paradev updated its platform with stacked pull request merging: a pull request in a stack now merges with the layers below it, and the merge button now reads Merge stack and lists each pull request before confirmation. Users can add a pull request or a whole stack to the base branch's merge queue. m:…5089
Every agent pane can now open its session in the other view, and cloud workspaces that run out of CPU or memory now show a dot in the sidebar. Claude Fable 5.2 is now available in the Claude model picker and in chat, and chat sessions now remember fast mode for each agent. m:…5089
Stacked PRs and merge queues are plumbing for a world where agents write code in layers and humans review the stack. The tooling is catching up to the fact that the unit of work is no longer one pull request at a time. m:…5089
Teams reorganize in opposite directions
Larridin surveyed engineering leaders about organizational changes for AI at LDX3 New York in September 2026. Over the past year, 40.8% of surveyed engineering leaders reorganized their teams around products or outcomes, while 29.6% had not changed their organization at all. m:…7679
The rest split in opposite directions: 23.5% widened their manager spans and 23.5% narrowed them, 19.4% changed the mix of senior and junior engineers, and 15.3% introduced roles focused on verification, agent platforms, or AI enablement. m:…7679
Almost 90% of directors reported some form of agentic coding in their interview loops. Agreement on technical foundations for AI is arriving faster than agreement on the organization that should run them, which is why copying another company's org chart is a gamble: the problems companies face differ, and so do the answers. m:…7679
Judgment is the differentiator
The Every Media post argues that human judgment regarding quality is the primary differentiator when all users have access to the same models. When everyone has the same models, what sets a team apart is its taste: what good looks like, and which output to ship. m:…6758
Mike Taylor proposes Level 9 (Simulation) and Level 10 (Multiverse) as advanced stages of AI adoption. Goodfire research shows that most neurons in an open-weight model perform legible jobs, and MIT researchers demonstrated that a frontier model can rewrite a smaller model's weights without a training run. m:…6758
The tools are getting better at long, ambitious builds and at customization. Fable 5.2 is effective for long, ambitious builds but is slow and expensive, Monologue is three times faster and cheaper than a trillion-parameter model, and Gemini 4 Carbon is a stronger coder than Gemini 4 Argon but lacks judgment on when to stop. m:…6758
What is still unresolved
- Which org model wins, or whether the split persists
- Whether weight-space customization moves from research into products teams can buy
- Whether skills libraries become shared across companies or stay internal
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.