Blog

Running many AI coding agents on one codebase.

What breaks when agents and people share a repository, and how to keep the work moving with a human in the loop. Written from building SyncStaff with SyncStaff.

A nine-item team checklist for Claude Code: agents run the tests, plan first, CLAUDE.md under 200 lines, shared settings, ask rules, a worktree per session, fresh-context review, watch the cost, and check in before edits.

Claude Code Best Practices for Teams Running Many Agents

Anthropic's Claude Code best practices, from verification to context management, plus the team layer: shared config, human checkpoints and coordination.

Jake6 min read
Claude Code hooks at fixed points: SessionStart registers, PreToolUse checks the file lease before each edit, and a fail-closed check refuses when unsure.

Claude Code Hooks: Guardrails That Always Run

How Claude Code hooks work, the events worth knowing, five hooks every team should commit, and the fail-closed preflight hook we run on every agent.

Jake7 min read
Plan mode as five steps: Shift+Tab turns it on, Claude writes a plan listing its files, Ctrl+G edits the plan, you approve it, and the build starts with those files claimed so other agents can see them.

Claude Code Plan Mode: Plan, Approve, Then Build

How Claude Code plan mode works: entering it, approving and editing plans, when to skip it, how it behaves in agent teams, and using plans to claim files.

Jake6 min read
The Claude Code sandbox as a boundary inside your machine: shell commands like npm test and git commit run inside, writing only to the project and temp and reaching only allowed domains, while file tools, hooks, MCP servers and ! commands sit outside it.

Claude Code Sandbox: What It Protects and What It Doesn't

How the Claude Code sandbox works: what it restricts, how to enable it, what runs outside it, its documented limits, and how it fits with permissions.

Jake7 min read
A tmux screen split into four panes: Claude Code sessions in their own worktrees for api and web, a Codex session for review, and the three ~/.tmux.conf lines that make Claude Code work inside tmux.

Claude Code in tmux: A Terminal Setup for Many Agents

The best terminal for Claude Code, the three tmux lines that fix Shift+Enter and notifications, split panes for agent teams, and a layout for many agents.

Jake6 min read
A receipt for Claude Code per developer: about $13 per active day, 90% of users under $30 a day, $150-250 a month in enterprise, roughly 7x tokens for agent teams in plan mode, and rework from collisions paid twice.

Claude Code Token Cost: What a Team Really Spends

Claude Code token cost explained: Anthropic's per-developer averages, how to check usage, billing by plan, where tokens go, and the waste parallel agents add.

Jake7 min read
A folder tree: the main checkout, where session edits are blocked, and .claude/worktrees/ holding feature-auth and bugfix-123, each on its own worktree- branch, plus a .worktreeinclude file for ignored files.

Claude Code Worktrees: Run Parallel Sessions Safely

Run parallel Claude Code sessions with --worktree: setup, how strictly isolation is enforced, cleanup and resume, and why conflicts still meet at merge.

Jake6 min read
Tiles for parallel agent tools: Claude Squad, Conductor, claude --worktree, claude agents and codex --worktree. A sixth tile shows the gap: your teammate's agents are visible to none of them.

Claude Squad, Conductor and Tools for Parallel Agents

Claude Squad, Conductor and the built-in worktree features compared for running parallel coding agents, and the problem none of them is built to solve.

Jake7 min read
A Claude Code session using the Codex plugin: /codex:review, /codex:adversarial-review, /codex:rescue and /codex:setup --enable-review-gate, each with what it does.

Codex Plugin for Claude Code: Two Vendors, One Workflow

How OpenAI's Codex plugin for Claude Code works: install, review and rescue commands, the review gate, and what it can't coordinate on a team.

Jake6 min read
A git commit graph: main checked out in ~/project, with branches feature-a and fix-issue-456 each checked out in their own linked worktree folder. Checking out one branch in two worktrees is refused.

Git Worktree vs Branch: When You Need Both

Git worktree vs branch explained: what each is, the commands, the one-branch-per-worktree rule, and why parallel AI agents need worktrees, not just branches.

Jake7 min read
OpenCode versus Claude Code: OpenCode reads AGENTS.md first, allows most actions by default, and leaves parallel worktrees to you; Claude Code reads CLAUDE.md first, starts in auto mode with a classifier, and has claude --worktree.

OpenCode vs Claude Code: We Tested Both on One Repo

OpenCode vs Claude Code compared hands-on: instruction files, default permissions, parallel sessions on one file, and how to run both on one repository.

Jake6 min read
METR 2025: experienced developers expected AI to make them 24% faster and believed afterwards it had made them 20% faster, but measured task time was 19% slower.

The AI Productivity Paradox: Why Teams Don't Get Faster

Developers feel faster with AI, but teams don't ship proportionally more. What METR and DORA found, and why coordination is the real bottleneck.

Jake6 min read
Diagram: a Claude Code team lead and two teammates around one file, api/routes.ts. Task claims are locked; api-dev owns the file and its edit goes through; frontend-dev's edit is refused by a hook.

Claude Code Agent Teams: What Breaks When Teammates Collide

Set up Claude Code agent teams, keep teammates from overwriting each other's files with an ownership hook, and see where a team's coordination stops.

Jake11 min read
Three columns of Claude Code permission rules: deny (Read of .env, force push), ask (git push, billing edits, migrations) and allow (npm test, git status). Ask rules reach a person even in auto mode.

Claude Code Permissions for Teams: Modes, Rules, Limits

The six Claude Code permission modes, what auto mode's classifier catches and misses, and how to set permissions so the right decisions reach a person.

Jake7 min read
Four stacked Claude Code settings files in order of precedence: managed-settings.json, .claude/settings.local.json, the committed .claude/settings.json, and ~/.claude/settings.json. Deny rules always win.

Claude Code Settings for Teams: What to Commit and Enforce

Where each Claude Code setting belongs, how the files combine, what a team should commit, and the few settings worth enforcing with managed settings.

Jake7 min read
Subagents versus agent teams: subagents run inside one session and report to the main agent, and in our same-line test one edit was refused; agent teams are separate instances on a shared task list, and same-file edits overwrite.

Claude Code Subagents vs Agent Teams: Which to Use

How Claude Code subagents work, how they differ from agent teams, and what happened when two subagents edited the same line, shared and in worktrees.

Jake8 min read
A CLAUDE.md file open in an editor: line one imports @AGENTS.md, followed by build and test commands and where things live, so Claude Code, Codex and Gemini share one set of rules.

CLAUDE.md Best Practices for Teams Running Many Agents

What belongs in a team CLAUDE.md, what to keep out, the AGENTS.md trap for mixed-tool teams, and why your most important rules also need a hook.

Jake7 min read
A terminal running two codex exec --worktree sessions on the same line: git worktree list shows two detached worktrees, and git merge agent-b ends in a merge conflict in greet.js.

Codex Worktrees: Parallel Agents Without Overwrites

We ran two Codex agents on the same line with --worktree. See where worktrees live, why the agents couldn't commit, and where the conflict surfaced.

Jake7 min read
568 decisions SyncStaff agents couldn't make alone, from August 3 to October 18, 2026: 356 answered by the founders and 161 resolved by the Chief of Staff.

Human in the Loop for AI Agents: 568 Escalations Later

How to keep a person in the loop for AI coding agents without drowning in prompts: decide by importance, route to chat, and check what you delegate.

Jake6 min read
Powered by Sync Chief of Staff