systems around the model

governance agent-infrastructure artifact-workflows

systems around the model

self.md radar — 2026-04-26

The model is no longer the interesting unit by itself; the action moved into the systems wrapped around it — swarm coordination, native artifact workflows, and the brittle plumbing that decides what context an agent actually sees.

The lead is a Science policy-forum warning that AI personas can manufacture consensus at civic scale. From there, one workflow turns an AI IDE into a native PowerPoint factory, and a three-example context-control cluster (RTK, SkillNote, Cursor’s AGENTS.md bug) makes the infrastructure point concrete.

1. synthetic publics arrive as a governance problem

sources:

what happened: A Science policy forum piece warns that large coordinated groups of AI-generated personas can imitate human behavior online and shape public opinion at scale. The accompanying summaries describe a single operator running vast networks of AI voices, executing millions of persuasion experiments in parallel, and manufacturing the appearance of broad agreement. The framing treats this as a democracy and trust-surface problem, distinct from one-off deepfakes.

why this matters: The countermeasure shifts from “spot the fake asset” to authenticated identity and verifiable provenance on the publics themselves. Media literacy alone does not scale against synthetic crowds.

2. PPT Master makes the AI IDE a native deck factory

sources:

what happened: PPT Master takes a PDF, DOCX, URL, or Markdown file and returns a natively editable PowerPoint with real shapes, text boxes, and charts — not flattened images. It runs as a workflow inside Claude Code, Cursor, VS Code + Copilot, or Codebuddy, with the README citing as low as $0.08 per deck, a local pipeline except for model calls, and parity across Claude, GPT, Gemini, and Kimi. The distinctive bit is the output artifact: a real PPTX, editable in Office, not a screenshot of one.

why this matters: AI IDEs are starting to behave like runtimes for business artifacts, not coding chat shells with extras. The unit of value is the editable file the team actually opens on Monday.

3. context plumbing becomes its own toolchain

sources:

what happened: RTK ships a single-binary Rust CLI proxy that filters and compresses command output before it reaches the model, claiming 60–90% token reduction across 100+ supported commands with under 10ms overhead, installed via a hook that rewrites bash invocations to RTK equivalents. SkillNote targets a different leak: Claude Code allots roughly 8,000 characters across active skill descriptions, beyond which skills get silently truncated, so it adds collections, live sync, and hook-based preservation across compaction and subagent spawns. And on Cursor’s forum, maintainers confirmed AGENTS.md was being routed into agent_requestable_workspace_rules instead of always_applied_workspace_rules, called it a bug tied to a broader alwaysApply regression, and pointed users to @AGENTS.md as a temporary workaround.

why this matters: Operators have stopped treating context as prompt text and started treating it as infrastructure — compressed, routed, synced, and audited. When the routing layer silently misfiles your rules, the model is doing exactly what you told it to; you just did not know what you told it.

left on the table