Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

tool discovery, skill bills, repo context

agent work is getting pushed through a colder filter: tool discovery, skill overhead, and repo context now need receipts instead of vibes.

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

self.md radar — 2026-07-22

agent work is getting pushed through a colder kind of filter: not whether the model sounds competent, but whether the surrounding system can find the right tool, carry the right context, and prove the extra instructions helped.

three objects are worth keeping open: a DNS-shaped proposal for agent tool discovery, a skills benchmark with an annoying answer, and JetBrains turning repo intelligence into a measured product surface. the day is not glamorous. good. glamorous is how agents end up with root access and a vibes-based procurement memo.

1. ToolDNS puts tool discovery back into plumbing

sources:

what happened: A new arXiv paper proposed ToolDNS, a way for autonomous agents to discover tools through DNS-style hierarchy instead of another centralized registry. the paper says its benchmark spans 33,688 real-world tools across MCP, A2A, REST, and Skill protocols. on that dataset, ToolDNS cuts the per-query search space by 95.26% while matching state-of-the-art retrieval accuracy, using DNS concepts like partially unfolded names, EDNS0 intent payloads, and logical subdomains.

why this matters: if agents are going to call millions of tools, discovery becomes a governance surface, not a search box. the self.md read is simple: the next personal AI stack needs boring addressability before it needs another shiny tool marketplace.

2. skills got a mixed grade

sources:

what happened: OrcaBot Labs tested five public skill collections with Codex 5.5 against SWE-bench Pro and SlopCodeBench. the result was not the comforting “skills always help” story. skills helped more on difficult repository tasks where tests, nearby patterns, and existing regressions gave the agent something to latch onto, but they often hurt SlopCodeBench’s smaller greenfield tasks, and most collections cost more to run than baseline.

why this matters: skills are not magic prompt vitamins; they are context with a bill. that is actually useful, because it moves agent craft from religion into measurement: use procedural memory where the repo can repay it, and stop stuffing every job with laminated wisdom.

3. JetBrains sells repo context with receipts

sources:

what happened: JetBrains introduced JetBrains Context, a repository-intelligence layer for coding agents. the launch says agents can use it for semantic code search, dependency/API validation, remote-repository lookup, and cross-org code discovery rather than relying only on files already checked out locally. JetBrains says it tested the system on 205 open-source SWE-bench tasks, 175 production-monorepo tasks, and 1,953 code-localization tasks, with agent turns down by up to 68%, latency down by up to 59%, and execution cost down by up to 48%.

why this matters: this is a repeat theme with a new object layer: repo context is no longer just a clever graph demo, it is being packaged as infrastructure with benchmark claims and IDE distribution. personal AI will hit the same problem at smaller scale: the agent cannot be “smart about my world” if it keeps relearning where the bodies, invoices, scripts, and half-finished notes are buried.

  • SAAG — breaks agent-call evaluation into registry conformance, structural completeness, and argument grounding instead of one fake-clean exact-match score.
  • Libretto Browser Tools SDK — gives agents six browser tools around Playwright, with a compact snapshot plus executable browser step instead of raw-page soup.
  • Nativ — local Mac model runner with telemetry, partner model picks, and integrations for Codex, Claude Code, Hermes, OpenCode, and Pi.
  • ai-lint — lints CLAUDE.md, AGENTS.md, SKILL.md, .cursorrules, and related prompt files for conflicts, duplication, missing frontmatter, and bloat.

left on the table

  • LifeOS has huge self.md overlap and fresh activity, but the current signal is still too broad for a main slot: more manifesto than operational delta.
  • LightMem2 was tempting because its README claims 95.7% fewer input tokens on OpenClaw continuous mode, but memory-cost stories have been heavy all week; support-grade unless the benchmark paper becomes the object.
  • Browser Tools SDK stayed in support because Monday already covered computer-use driver layers; useful, not different enough for main.
  • Agentic Engineering Protocol and ai-lint reinforced the “agent instructions need verification” pattern, but after yesterday’s planner-injection lead they are footnotes, not the lead car.
  • the JetBrains context repeat is allowed because it brings a first-party launch plus benchmark claims, not just another repo-context wrapper.
agents tooling coding-agents personal-ai