I spent this week cleaning up after agents, so the tools that caught my eye were the ones that make agents easier to trust and cheaper to run. Two others made it in just because they’re fast and I like fast.
1. Give your agents one shared memory they can branch
Omnigraph is a graph database for teams running several agents on the same knowledge: a typed schema enforced on every write, graph, vector and full-text search in one query, and everything stored on S3-compatible storage in the open Lance format.
2. Let your coding agent paint with real brushes
Klad is a painting and animation app you share with your agent: you draw with a pen, Claude Code or Codex paints through the klad CLI with the same brushes, and you both see each other's strokes live. There's no image model in it, so if one leaf looks off you fix that leaf instead of rerolling the whole picture. It's macOS for now and source-available, free for personal use and for companies under €500k revenue.
3. A database client that opens in a second
DBDelve is an open source DB client built in Rust on GPUI, Zed's UI framework. In the author's benchmark it launches in 1.1 s against DBeaver's 5.2 s and idles at 133 MB instead of 606 MB, and it covers Postgres, MySQL, SQLite, DuckDB, SQL Server, Snowflake, MongoDB, Redis and ClickHouse.
4. See what's actually eating your GPU memory
Task Manager tells you "15 GB used". gputaskman from Scott Hanselman tells you how much of that is model weights, KV cache, your browser or the compositor, per process, in a terminal UI. One uv tool install gputaskman and it runs on Windows, Linux, WSL and Apple Silicon. It's early and the model breakdowns are estimates, but it's the first tool I've seen that answers "why won't this model fit".
5. Teach a cheap model the job your expensive one figured out
AgentRun sits on top of Claude Agents, Codex, Pi or OpenCode and adds durable runs, a control plane and a "learning" step that carries what one run figured out into the next. The team built it after their code review bill went from $180 to $6,451 a month; their own reviewer on a cheap model now costs 18 cents a review. It's Apache 2.0 and self-hostable with an OpenRouter key. I'd test it on one repetitive job before trusting the self-improvement on anything bigger.
Building a devtool, or found one I should know about? Hit reply. I'd love to hear about it.
Jonathan (@jonathimer)