blog

Updates, deep-dives, and design notes.

The best AI agent orchestration tools for coding in 2026

A comparison of the tools that run multiple AI coding agents at once: Conductor, Vibe Kanban, Claude Squad, crystl, Cursor cloud agents, Sculptor, and doing it by hand with tmux and git worktrees.

Agent communication protocols: A2A, ACP, and MCP

A reference on the agent communication protocols people actually mean: A2A for agent to agent, MCP for agent to tool, and the two different things called ACP. What each one is, who governs it, and where the confusion comes from.

Running an agent team in tmux

A pane per role, send-keys to dispatch, a shared file as the channel, and a polling loop for idle. The real wiring for a tmux agent team, and the four guarantees it doesn't have.

How do you get AI agents to talk to each other?

You have two coding agents open and neither can see the other. The four real options for agent to agent communication, worst to best: a shared file, a git branch, tmux send-keys, and a coordination layer. What each one guarantees and what it doesn't.

Claude Code Subagents: Best Practices

Scope a subagent to one question, state the return shape, and never delegate work you'd want to watch. Concrete examples of subagent tasks that work and ones that waste your tokens.

Cursor background agents vs crystl

Cursor's background agents, now called cloud agents, run Cursor's own agent in isolated cloud VMs and hand back a pull request. crystl runs any CLI agent locally across all your projects. What each one is actually for.

Git: Cannot Delete Branch Used by a Worktree

Git refuses to delete a branch that a worktree has checked out, even with -D. Find the worktree with git worktree list, then remove it or prune it, then delete the branch.

git worktree add: Every Form of the Command

git worktree add with a new branch, an existing branch, a remote branch, a commit, or a detached HEAD. Every flag that matters, with real examples and the errors you'll hit.

Git Worktree vs Branch (vs Multiple Clones)

A branch switches one working copy in place. A worktree gives you a second working copy on the same history. A clone gives you a whole separate repo. When to reach for each.

A Git Worktree Workflow That Stays Tidy

Where to put worktrees, how to name them, what to do about node_modules and .env, and when to remove them. The habits that keep a worktree workflow from turning into folder sprawl.

Git Worktrees in Cursor, Zed, and IntelliJ

How each editor handles git worktrees: Cursor's agent worktrees, Zed's worktree picker, IntelliJ's Worktrees tab, and the one thing that works everywhere.

How Many Agents Can You Actually Run at Once?

The ceiling on parallel agents isn't your machine. It's tokens, rate limits, and mostly your own review capacity. What actually binds first, why more agents past a point ships less code, and how to pick a number.

How to manage multiple projects when every project has an agent in it

Managing several projects used to be about your own time. When each project has a coding agent working in it, the constraint moves to your attention. Six rules that hold up: per-project context, written state, contained blast radius, batched review.

Agent orchestration, when what you actually have is too many tabs open

Agent orchestration is a term from agent-building frameworks: the code that decides which agent runs when, inside a program you write. If you searched it because you have nine terminal tabs and don't know which one needs you, that's a different problem with a different fix.

How to Remove a Git Worktree

git worktree remove deletes the directory and its metadata. git worktree prune cleans up after directories you deleted by hand. Neither one deletes the branch. Here's the full cleanup.

Can You Run Multiple Instances of Claude Code?

Yes. Claude Code has no cap on concurrent instances. What two sessions share, what they don't, why the same working directory is the thing that breaks, and where the real limit sits.

Skills vs Subagents in Claude Code

A skill loads instructions into the conversation you're already having. A subagent runs in a context you can't see and hands back a summary. The real difference, the file formats, and when each one is wrong.

The tmux setup people build for parallel agents, and what it can't do

Claude Code ships a --worktree --tmux flag that covers one agent. Everyone running several converges on the same hand-built setup instead, and here are the four moments it stops working.

Claude Code has GUIs now. Here's what one should actually do.

Claude Code has a desktop app, a web app, IDE extensions, and a pile of community GUIs. Almost all of them build the same thing: a prettier chat window. That solves the easy problem. Here is the hard one.

Warp vs. crystl: Cloud Agent Platform or Local Fleet Cockpit

Warp vs. crystl for running AI coding agents: where Warp's cloud, enterprise, and platform strengths differ from crystl's local, personal, own-your-costs approach.

How to Use Git Worktrees with Claude Code

Claude Code worktrees explained: the --worktree flag, the exact git commands, real parallel-agent workflows, the gotchas nobody warns you about, and how to automate the whole thing.

What the heck is a "footgun" and other things Claude says

A footgun is a feature that makes it dangerously easy to hurt yourself. Why Claude Code talks the way it does, plus a plain-English glossary of 20 pieces of developer slang it drops into its answers: footgun, yak shaving, load-bearing, technical debt and more.

Introducing roundtable agent orchestration

Most multi-agent orchestration hides the agents inside one controller's head. Roundtable orchestration, the model behind crystl quest, puts independent agents in a shared chat you can actually watch.

How to Optimize Claude Code's Context Window (Without Just Buying a Bigger One)

Practical tactics for stretching Claude Code's context window: what to compact, what to delegate to subagents, and which plugins and skills to disable before they eat your tokens.

Do You Need a Code Editor to Build with Claude Code?

Claude Code runs in the terminal, reads your files, and writes your code. So do you still need VS Code open? The pros and cons of both setups.

Voice, Loops, and the Slow Disappearance of the Command Line

Claude Code's March 2026 features hint at something bigger than incremental updates. The relationship between developer and tool is shifting.

Claude Can Use Your Computer Now. That's a Stranger Sentence Than It Sounds.

Anthropic's computer use feature lets Claude control your Mac: clicking, typing, opening apps. What that actually means depends on who you are.

The Agentic Coding Workflow Nobody Talks About: Managing the Agents Themselves

The bottleneck in agentic coding isn't AI capability. It's managing multitask workflows where multiple AI agents are working simultaneously.

Why Your Terminal Wasn't Built for AI Pair Programming (And What To Use Instead)

Terminals were designed for human-typed commands, not for supervising AI agents. AI pair programming needs a different kind of tool.

Claude Code in iTerm2 vs. crystl: What Changes When You Use a Purpose-Built Terminal

A fair comparison of running Claude Code in iTerm2 versus crystl, a terminal designed for agentic coding workflows on macOS.

From Terminal Chaos to Crystal Clarity: Migrating Your Claude Code Workflow

A step-by-step guide to organizing your Claude Code setup on macOS. Move from scattered terminal sessions to a structured workflow.

Test Suite Broken? Run Claude Code on the Fix While You Keep Building

How to multitask with parallel Claude Code sessions: fix a broken test suite in one session while continuing feature work in another.

How I Use Claude Code to Ship a PR Every Hour

A workflow for structuring parallel Claude Code tasks across sessions so you always have something ready to review and merge.

Managing Claude Code API Keys Across Multiple Projects

How to avoid leaking API keys across projects when using Claude Code, and a cleaner approach to per-project key management.

My Claude Code Workflow for Refactoring a Legacy Codebase

How I used Claude Code to refactor a legacy codebase by breaking the work into parallel tracks and merging incrementally.

Per-Project MCP Server Configs for Claude Code: Why and How

How to manage different MCP server configurations for each Claude Code project, and why a per-project setup prevents config headaches.

Organizing 10+ Claude Code Projects Without Drowning in Terminal Windows

How to multitask with Claude Code across many projects without losing track of sessions. The gems and shards mental model for terminal organization.

Running Claude Code in the Background: A Workflow Guide

How to run Claude Code agents in the background, check their status, and structure tasks for async AI workflows.

How to Monitor Multiple Claude Code Agents Without Losing Your Mind

Monitoring parallel Claude Code agents: how to tell which one is blocked on an approval, which finished, and which went off the rails, without clicking through every tab. What tmux surfaces, what it can't, and what agent observability actually needs.

The Tab-Hunting Problem: Finding Which Claude Code Session Needs Your Attention

Running 5+ Claude Code sessions? Why you can't find the one waiting for approval, and how to fix it.

Why Your Claude Code Agents Keep Overwriting Each Other's Files

Running multiple Claude Code agents on the same repo? Why they conflict and how to fix it with isolated worktrees.

Fix a Bug and Build a Feature at the Same Time with Claude Code

A walkthrough of running two Claude Code sessions in parallel, one fixing a bug and one building a feature, in separate git worktrees.

How to Run Two Claude Code Agents on the Same Repo Without Destroying Your Git History

Multitasking with two Claude Code agents on the same branch causes merge conflicts. The fix is git worktrees, and you can automate it.

The Case for a Persistent Claude Code Session Log

Why your Claude Code session history should be persistent and browsable, not trapped in terminal scrollback that disappears when you close the tab.

What Did Claude Do While I Was Gone? Reviewing Agent Work After the Fact

Kicked off a Claude Code task and stepped away? How to review what your agent did, what it changed, and what it asked while you were gone.

I Closed My Terminal and Lost Everything: How to Never Lose a Claude Code Session Again

Lost a Claude Code conversation after closing your terminal? Why it happens and how to preserve your session history.