Builders are tearing out last year's agent scaffolding: routers, sandbox VMs, graph builders
The lead
For six straight days this brief led with agents arriving somewhere new. Today the movement runs the other way: the scaffolding built for them over the past year is being pulled out and replaced. Atlassian rebuilt Rovo Chat around a single long-horizon reasoning engine, concluding that the router wiring its per-product agents together was itself the constraint. Dify shipped a sandboxed Linux agent runtime and a Skill Editor, leaving its visual workflow builder as the legacy layer. Rivet argues coding agents never needed a sandbox VM at all — its agentOS gives them a Linux-shaped terminal on WebAssembly and V8 isolates — and Vercel's Eve and Flue took the bet.
Nobody agrees on the direction. FuseBase made the opposite call: its new Flow product splits AI development into phases, slices and gates, on the thesis that one chat collapses under strategy, product and build work at once. Router out in one place, structure in at another — the shared premise is that the first architecture was wrong, not that it needs more features. MCP endpoints kept landing today too, at Trello, Zoho Sign, Canix and Wheelhouse. That has stopped being the story; it is plumbing now.
What moved
- Agents shipping trust, not capability. Retool's agent now asks clarifying questions and reads branches created outside Retool; Aha! added authentication and virtual user testing to Builder; n8n's release week carries no Features section at all, the fixes concentrated in its AI builder; LangGraph's 1.2 line is checkpoint and subgraph durability under graphs deeper than it was built for.
- Retrieval and modeling as rented layers. Sourcegraph launched Code Finder and Agentic Batch Changes aimed at agents rather than humans, claiming a cheaper model plus its MCP server beats a frontier model alone. Omni took AI semantic-model generation to GA — the labor BI vendors sell as craft. AWS keeps building AgentCore one layer below whichever framework a customer picks.
- Supply-chain enforcement moved to publish time. GitHub now scans npm packages as they publish, requires dual-use metadata, and holds suspicious Actions workflows on public repos for approval — gating in the pipes it owns rather than advising after the fact. Argo CD's 3.5 train carries supply-chain verification, Apache Pinot 1.5.1 was pure CVE cleanup, and mediasoup spent the quarter hardening the SFU much of the industry's video stack rests on.
- Releases with no agent in them. Gravity Forms 3.0 makes accessible markup the default across every form; Spree Commerce 5.6 puts a wholesale portal inside the same store serving retail; Moov now moves money on both US instant rails.
Sectors today
- Devtools — Rivet, Sourcegraph and incident.io treat the agent as the customer; Merge shipped connector maintenance.
- AI assistants — the day in miniature: Dify rewriting its runtime, LangGraph stabilising the last one.
- Development — GitHub and Argo CD moved verification earlier in the pipeline; Zed and Sanity turned the crank another week.
- Ecommerce — Spree's wholesale portal and SureMembers team accounts are the sector's only real news amid content mills.
- Video conferencing — Element Call let federation pick the SFU; JW Player built a short-form surface; Panopto went after corporate L&D.
- Collaboration — Skedda and Paperless-ngx are both in the unglamorous phase after a big swing.
- Design — Balsamiq is making editor primitives machine-legible before generation can be reliable; Frame.io and Air push AI into review.
- HR and recruiting — Workable is becoming a system of record agents can drive; Tanda wires scoring into rostering.
- Analytics — Omni, Holistics and Displayr are all making AI output legible to the analyst defending it.
- Finance — Intuit is teaching the bank feed to show its work; Moov and CloudZero ship steadily beneath it.
- Project management — Atlassian's rewrite dominates; Rize spent two months fixing the app around its agent-queryable data.
- CRM — Twenty is quietly becoming an app platform with a marketplace; the rest publishes listicles.
- Marketing — ContentStudio started tracking ad spend; Statusbrew added agency plumbing.
- Marketing automation — n8n's reliability triage, Mautic patching three branches at once, Buttondown turning archives into websites.
- Customer support — Respond.io is dropping the phone number as identity; Twilio's month is compliance plumbing.
- Communication — SimpleX Chat is inventing trust for public channels without identifiers; Synapse lets Rust eat its internals.
Watch tomorrow
Three threads have a visible next move. Dify called its agent runtime experimental, warning to expose it only to trusted users; shedding that label depends on the tenant-scoping work filling its recent patches. Atlassian's long-horizon engine runs in Rovo Chat alone; Jira and Confluence are the obvious next surfaces. Rivet's open question is pricing, unanswered in any entry: a runtime living inside someone else's backend still has to be metered somehow. One caveat on today's counts — Apache Superset's feed tracks its Helm chart, not the project, and Metricool, Printful and Qandle are SEO blogs. Their velocity is real; their activity is not.