Open Source Radar — September 23: the agent's missing plumbing
Today's board is the three layers an agent needs before anyone trusts it with real work: a document surface it can actually edit, a runtime that keeps it isolated and cheap to wake, and a way to hand it tools without handing it keys.
Univer (TypeScript, ~16,000 stars, Apache-2.0) — An office SDK that has spent three years being the open-source spreadsheet you embed in your own product, and has now repositioned itself, in its own README, as the Office Harness for AI Agents. Spreadsheets, documents, presentations, relational tables and boards share one canvas-based runtime, one formula engine and one API that behaves the same in the browser and on Node — which is the part that matters here: a server-side process can read and write the same workbook a human has open. The repo ships presets and a plugin architecture rather than a fixed app, and the family around it is where the agent story lives: a CLI for agents to create and deliver office files locally, a plugin for the DeepSeek Harness, one for OpenClaw, and a workspace where agent edits get reviewed. If your agent's deliverable is still a markdown blob the user pastes into Excel, this is the missing half.
Agent Substrate (Go, ~3,200 stars, Apache-2.0) — Google's take on the layer under the agent: explicitly not an SDK for building agents, a system for running them at scale. It maps a large set of mostly-idle actors onto a small pool of workers, suspends and resumes them with RAM and filesystem state intact, and claims sub-500-millisecond resume at more than 500 activations a second — the demo juggles 250 stateful actors across eight pods, roughly 30x oversubscription. The sandboxes are real ones, microVMs or gVisor, with kernel and network isolation, and Kubernetes handles provisioning underneath. It's framework-agnostic by design — LangChain, ADK, coding agents and sandboxed MCP servers all run as actors, and the CNCF project kagent is already building on it. The caveat is in the README, not buried: early development, not production-ready, APIs will change. Read it anyway, because this is the shape the agent-runtime race is converging on.
treg (Python, ~2,400 stars) — OpenRouter's model applied to tools instead of models: one base URL and one token, and the agent can call 3,000-plus catalogued endpoints across 60-plus providers — SEO and backlink data, people and company enrichment, ads, scraping, image and video generation — priced per call from a cent, with no vendor signup. The reason it exists is arithmetic: Semrush at $139 a month, Moz at $99, Crunchbase at $99 are subscriptions nobody buys for a single agent run, so treg holds the account and bills fractions of a cent instead. Your own keys always win over its pool and those calls are never metered, credentials are injected server-side so callers never hold them, and the same registry fronts your team's vendor CLIs and skill files. Two things to know before adopting: the licence is Apache-2.0 plus terms barring third-party hosted offerings without permission, and the free tier is a one-time dollar of credit. It runs hosted at treg.to, and self-hosting is supported.
Worth watching this week: whether the runtime and the tool registry — not the model — become where agent platforms actually compete.
Which layer of your agent stack do you still run yourself: the document surface, the sandbox, or the credentials? Tell us in the comments.
Sources: Univer (GitHub) · Univer SDK docs · Agent Substrate (GitHub) · Agent Substrate demo · treg (GitHub) · treg
