Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

agent work got useful where it slowed down

today’s useful agent work had friction in it: a correctable personal model, coding workflows that keep the human thinking, and tool catalogs that load late instead of eating context.

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

self.md radar — 2026-05-30

agent work got useful today where somebody added friction on purpose.

the personal-agent story was not another assistant wrapper; it was a visible model of the person. the coding story was not “go faster”; it was how to keep your own brain in the loop. and the tool-protocol fight got boring in the right way: big tool catalogs now need loading policy.

1. the personal model became the product

sources:

what happened: Elephant Agent, a repo created on May 15 and active today with 562 GitHub stars at verification, frames itself as a “Personal-Model First Self Evolving AI Agent.” the README says the agent starts from the person, not the task, and grows a correctable Personal Model across Identity, World, Pulse, and Journey. its paper page calls the project “Agency-First Personal AI” and centers personal-model-first evolution for judgment, evidence, questions, and growth.

why this matters: a personal AI with a visible user model is closer to an operating system than a chat tab. the catch is maintenance: if Identity, World, Pulse, and Journey become steering state, they need edits, provenance, expiry, and a way to say “no, that was a mood, not a truth.”

2. coding agents got a speed bump

sources:

what happened: Vicki Boykis wrote that agentic code generation can leave her with “the outward signs of having written code” but not the internal process that makes the code legible in her own head. her repair list is deliberately slow: write the first implementation herself, ask the agent questions instead of taking patches, wait 20 minutes before calling it in, read papers and books, and reimplement fundamentals. a tiny Cartographer skill pushes the same move into Claude Code: build the problem-theory of a domain before designing or coding.

why this matters: the useful agent workflow may be the one that makes the human do a little more work first. if delegation skips the part where the operator forms a theory, the repo gets code and the person gets fog.

3. tool catalogs learned to load late

sources:

what happened: Quandri measured its real MCP stack and said four MCP servers consumed 10.5% of the context window from tool definitions alone, with Linear taking more than 12,800 tokens across 42 tool definitions. the same post says a Linear CLI lookup used about 65x fewer tokens than the MCP path, then adds the important update: Claude Code Tool Search with Deferred Loading now loads MCP schemas on demand and reduces the context-bloat problem by 85%+ for current users. Claude’s docs say tool search is enabled by default, only tool names load at session start, and only tools Claude actually uses enter context; OpenAI’s docs expose the same shape through tool_search and defer_loading: true for functions and MCP servers on gpt-5.4+.

why this matters: the argument is not really “MCP is dead.” it is that the context window is now a budget with lazy-loading rules. personal agents need tool access, but dumping every menu onto the table at startup is just another way to spend attention before the work begins.

  • EveryInc compound engineering plugin — a cross-editor plugin pack for Claude Code, Codex, Cursor, and more; useful, but too close to this week’s plugin-folder story for a main slot.
  • Cursor plugins — official plugin manifests and marketplace plumbing; another sign that agent work is turning into installable packages with names and permissions.
  • VTCode — a Rust terminal coding agent with skills, subagents, hooks, sandboxing, MCP, and ACP; fresh enough to watch, not fresh enough to lead.
  • Sverklo — local-first repo memory for coding agents with 37 tools, git-pinned decisions, and no code upload; a nice footnote to the Personal Model signal.
  • PingCAP mem9 — persistent memory moved from API surface to ingestion, ranking, evaluation, shared spaces, and inspection work.

left on the table

personal-ai agent-workflows agent-memory mcp coding-agents