Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

runtime, files, facts

Bastion isolated coding agents in VMs, Google proposed a Markdown knowledge format, and the Census noise fight turned public data into a control surface.

Published Updated Read 4 min
■ [!] ON THIS PAGE ▼

self.md radar — 2026-06-14

the useful objects today were not smarter chats. they were places to run agents, files to feed them, and rules about what counts as trustworthy data.

Bastion and Fleet made coding agents look like a small batch system. Google put agent-readable knowledge into Markdown directories with YAML frontmatter. the Census noise fight was the ugly governance version of the same problem: once machines consume “official” facts at scale, the provenance layer is not neutral.

1. parallel agents need walls

sources:

what happened: Bastion launched around a blunt promise: virtual computers for coding agents, with each agent running in its own VM so parallel work does not step on the same runtime. Fleet took the home-lab version of the same problem: a supervisor, task queue, background daemon, Beads database, and adapters for Claude Code, Codex, opencode, and other coder CLIs. Xiaomi’s MiMo Code added the long-horizon side, open-sourcing an OpenCode-based terminal agent meant to keep state across dozens or hundreds of execution steps with a Markdown project-memory file.

why this matters: The serious coding-agent problem has left the prompt box: where the workers run, who isolates them, what keeps the queue fed, and whether the hardware is busy enough to justify the bill.

2. Google put knowledge in plain files

sources:

what happened: Google Cloud published Open Knowledge Format v0.1 as a directory of Markdown files with YAML frontmatter, not a new database or runtime. the spec requires a type field, recommends fields like title, description, resource, tags, and timestamp, and reserves index.md for progressive disclosure plus log.md for update history. the repo also includes Knowledge Catalog support and sample tooling, while carefully saying the project is not an official Google product.

why this matters: This is close to the self.md bet in miniature: agents need portable knowledge bundles they can inspect, cite, diff, and move between systems. the funny part is how little magic is required when the useful contract is just files, metadata, links, and restraint.

3. statistical noise became a control surface

sources:

what happened: Ted Desfontaines traced a Commerce Department order that bans “noise infusion” from statistical products published by the Census Bureau and the Bureau of Economic Analysis. noise infusion is one of the privacy techniques used when agencies publish aggregate numbers from confidential datasets. the article’s warning is simple: removing noise can make tables feel cleaner while making re-identification and linkage attacks easier.

why this matters: Personal agents will happily treat public statistics as ground truth. if the official data layer changes its privacy guarantees, the downstream machine does not just inherit better numbers; it inherits a political decision about whose privacy can be shaved off to make the spreadsheet look crisp.

  • QuantmLayer — a kernel-enforced containment cell for coding agents, with a learning mode that records what the agent actually needed before tightening the profile.
  • Eywa — the strongest memory paper in today’s pile because it starts with immutable source evidence before deriving long-term facts.
  • Open Knowledge Format samples — useful because the examples show OKF as data-team plumbing, not just another Markdown manifesto.
  • Loopy — a meta-loop wrapper for Claude Code and Codex sessions; too small for a main slot, but very much part of the “agent as worker process” pile.

left on the table

  • Fable export-control follow-ups stayed out because yesterday already used the Anthropic/Fable shutdown as the access-control story.
  • local-first memory MCP tools had the right smell, but memory has been a main thread all week; Eywa carried the cluster as a supporting link instead.
  • QuantmLayer was real but too close to last week’s containment and prompt-injection coverage to take a numbered slot.
  • GLM 5.2 looked like model-release noise in this cache, not a self.md-level operational shift.
  • Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
  • AI coding assistants — compare coding workbenches by review surface, permissions, cost, logs, and escape hatches
  • Best MCP servers — connect files, browsers, memory, search, and workflow tools without turning the stack into soup
agent-ops knowledge-management data-governance