Practical notes on local-first AI agent workflows.
Writing on controlling Claude Code, Codex, and OpenCode from anywhere, keeping your repositories on your own machine, and building calmer multi-agent development habits.
More field notes
90 postsDetect Prompt Regressions in AI Coding Agents
Compare old and new prompts with fixed tasks so Claude Code and Codex do not drift into broader or noisier behavior.
Shift Handoff for AI Agents: A Developer's Playbook
Hand off active Claude Code and Codex work between people without losing context, ownership, or review history.
How to Build an AI Agent Escalation Policy
Create a simple rule set for when an agent should continue, pause, ask for approval, or hand work back to a human.
How to Route AI Agent Work by Task Severity
Route low, medium, and high-risk Claude Code and Codex work to different approval paths and supervision levels.
How to Handle Security-Sensitive AI Agent Work Safely
Keep Claude Code and Codex away from risky edits unless the task, approval path, and diff are tightly scoped.
How to Write a Destructive Command Policy for AI Agents
Define when rm, reset, migration, and deployment commands need human approval so Claude Code and Codex stay bounded.
How to Run a Docs-Only AI Agent Workflow
Use Claude Code or Codex for docs-only tasks with tight scope, fast review, and fewer unnecessary approvals.
How to Roll Out Switchboard Gradually Without Automating Everything
Introduce Switchboard with a phased rollout that starts small, protects review quality, and expands only after routes and issues are proven.
Flaky Test Fixes With AI Agents: A Safe Debugging Workflow
Separate real regressions from flaky tests, then use Claude Code or Codex to repair the test or the code with a tighter review loop.
How to Supervise a Long-Running Refactor With AI Agents
Keep a Claude Code or Codex refactor on track with checkpoints, diffs, approvals, and explicit stop points.
Turn a Bug Report Into an AI Agent Runbook
Convert a vague issue into a bounded Claude Code or Codex run with scope, checks, and a clear stop condition.
Why a Browser Control Surface Beats Remote Desktop for Claude Code and Codex
Compare browser-based agent control with remote desktop when supervising local Claude Code and Codex sessions from phones or laptops.
How to Run Overnight Bug-Fix Queues with Switchboard
Use Switchboard to keep a bounded Linear bug queue moving overnight while preserving review gates, routing discipline, and pull request quality.
How to Structure One Junction Account for a Small Team
Plan a small-team Junction setup with clear daemon ownership, provider auth boundaries, plan limits, review habits, and Switchboard timing.
How to Triage AI Agent Runs During Your Commute
Use short commute check-ins to review Claude Code and Codex progress, block risky work, and keep useful runs moving.
AI Coding Agent Incident Response: A Practical Playbook
Respond to production incidents with Claude Code or Codex using tighter scope, live monitoring, and a clean handoff from diagnosis to patch.
How to Turn Handoff Notes into Agent Instructions
Convert repeated human handoff notes into reusable Claude Code and Codex instructions, templates, and review guidance inside Junction.
How to Prevent Claude Code and Codex Cross-Talk on the Same Repo
Run Claude Code and Codex on the same repository without mixing branches, prompts, context, approvals, or review paths.
How to Run an On-Call AI Agent Workflow
Build a calm on-call routine for Claude Code and Codex with scoped approvals, notifications, and mobile review.
How to Recover from a Bad Branch or Workspace Drift
Recover from AI agent branch drift by stopping the run, preserving evidence, separating changes, and restarting from a reviewable state.
How to Keep One Junction Account Clean When Machines Come and Go
Keep your Junction account understandable as laptops, workstations, and servers are added, retired, renamed, or repurposed.
How to Run AI Coding Agents During a Release Freeze
Use a release-freeze policy to keep Claude Code and Codex moving on safe work while protecting release-critical branches and approvals.
How to Keep Long-Running Agent Work Unblocked Across Time Zones
Build a time-zone handoff routine for long-running Claude Code and Codex work using session state, notes, notifications, and review gates.
How to Review Agent PRs on Mobile Without Slowing Yourself Down
Use mobile review for fast triage of AI agent pull requests while saving risky diffs, broad changes, and final merges for desktop.
How to Build an After-Hours Approval Routine for Agent Runs
Create a safer after-hours routine for Claude Code and Codex approvals using notifications, risk checks, and mobile review discipline.
How to Use Switchboard Analytics to Find Automation Friction
Use Switchboard run analytics and activity patterns to spot bad routes, unclear issues, flaky repos, and automation that needs review.
How to Tell Whether a Repository Is Ready for Switchboard
Evaluate tests, branch hygiene, instructions, review paths, and daemon setup before routing a repository into Switchboard automation.
How to Design Linear Statuses for Agent Automation
Shape Linear statuses so Switchboard automation has clear pickup, running, review, blocked, and done states for AI agent work.
How to Decide Which Issues Belong on Switchboard
Use a practical selection rubric to decide which Linear issues are ready for Switchboard automation and which should stay manual.
How to Organize Multiple Repositories in One Junction Workspace
Keep multiple repositories understandable in Junction with clear working directories, session names, Git state, and review habits.
How to Run an AI Agent Hotfix Queue
Build a tight hotfix queue for Claude Code and Codex so urgent fixes get fast attention without turning into an uncontrolled backlog.
How to Route Claude Code and Codex Jobs to Different Machines
Use task shape, repo location, credentials, and machine availability to decide where Claude Code or Codex should run.
How to Separate Personal and Work Daemons in Junction
Keep personal projects and work repositories cleanly separated with daemon boundaries, provider auth, Git identity, and plan choices.
When to Put Claude Code or Codex on a Headless VPS
Decide when a headless VPS is the right execution host for Claude Code or Codex, and when your laptop should stay the agent machine.
Set a Session Retention Policy for AI Agents
Keep useful Claude Code and Codex history without letting old sessions, branches, and notes take over the active surface.
Use a Dedicated Agent Box for Claude Code and Codex
Build a cleaner local AI agent workflow by running Claude Code and Codex on one stable machine while controlling them from any browser.
How to Clean Up Branches After AI Agent Runs
Decide when to merge, delete, prune, or keep agent branches so worktree cleanup stays reviewable and your repo stays readable.
How to Search AI Agent Session History for Useful Context
Find the right prior Claude Code or Codex run faster by searching sessions for branch, repo, outcome, and failure mode.
Use Agent Transcripts as an Audit Trail
Keep Claude Code and Codex sessions reviewable by capturing output, approvals, diffs, and the reason a run changed direction.
How to Write AI Coding Agent Runbooks That Actually Help
Turn common Claude Code and Codex tasks into short repo-local runbooks that reduce drift and make supervision easier.
How to Name AI Coding Agent Sessions So They Stay Findable
A naming scheme for Claude Code and Codex sessions that makes search, handoff, and cleanup faster across devices and worktrees.
How to Route Human Approvals for AI Agent Work
Build a simple approver map so Claude Code and Codex ask the right person at the right time for risky changes.
How to Triage Morning Agent Runs Before You Start Work
Use a repeatable morning pass to check Claude Code and Codex runs, clear blockers, and decide what deserves deeper review.
Agent Session Taxonomy: A Practical Way to Label and Search Runs
A practical taxonomy for naming AI agent sessions by intent, risk, and lifecycle so you can search, route, and archive them later.
Automation Candidate Grooming: How to Pick the Right AI Work
A practical filter for deciding which agent tasks are worth automating, and which should stay manual because the blast radius is too wide.
One Agent, One Task: The Discipline That Keeps Runs Reviewable
A simple rule for keeping AI agent runs reviewable: one run, one outcome, one branch, and one stop point.
Parallel Agent Experimentation: Compare Claude Code and Codex Fairly
How to run the same task with Claude Code and Codex in parallel so you can compare diffs, risk, and review effort without guesswork.
Stop Before Shared State Damage: A Better Rule for Agent Runs
A stop rule for AI agent runs that touch shared branches, migrations, or other repo state before a small mistake becomes expensive.
Weekly Agent Ops Checklist: Keep Claude Code and Codex Tidy
A weekly sweep for cleaning up runs, branches, notifications, and daemon state before small bits of drift become busywork.
Configure Switchboard Route Settings for Team and Repo Routing
Learn how Switchboard route settings map Linear teams and repositories to daemons, agents, models, branches, and automation rules.
Send Conflict-Fix Instructions Back to Agents
Turn merge conflicts and branch drift into a focused follow-up task for Claude Code or Codex instead of manually restarting work.
Use File Mentions and Image Attachments to Give Agents Better Context
Give Claude Code and Codex better starting context with file mentions, screenshots, and focused prompts in Junction.
Use Plan Mode Before Edits Start
Use plan mode in Junction to make Claude Code and Codex explain risky work before edits, commands, and pull requests begin.
Archive AI Agent Sessions Without Losing History
Keep active Claude Code and Codex chats manageable while preserving transcripts, review context, and useful handoff details.
Choose Models Per Chat for AI Coding Agents
Learn why per-chat model selection helps match Claude Code and Codex runs to task risk, cost, speed, and review expectations.
Keep Claude Code and Codex Auth Separate on Each Daemon
Learn how daemon-scoped Claude Code, Codex, GitHub, and model settings keep local AI coding workflows easier to route and debug.
Use CLAUDE.md and AGENTS.md to Steer Local Agent Runs
Write durable repository instructions that help Claude Code and Codex follow project rules while Junction keeps execution local.
Resume AI Agent Sessions Across Devices
Learn how to keep Claude Code and Codex context available when you move between desktop, browser, phone, and archived history.
Stop and Steer Live AI Agent Runs Before They Drift
Learn when to stop, redirect, or continue Claude Code and Codex runs before a small mistake turns into a large diff.
Use the Terminal and File Browser as an Agent Control Surface
See how terminal output, workspace files, diffs, and agent transcripts work together when supervising Claude Code and Codex.
Plan Around AI Coding Agent Rate Limits
Learn how to plan Claude Code and Codex workflows around usage windows, quotas, local telemetry, and review timing without guessing.
Use Branch Suggestions to Keep Agent Runs Reviewable
Keep Claude Code and Codex work reviewable by starting each local agent run on the right branch, worktree, and pull request path.
Use Live Output Streaming to Catch Agent Problems Early
Learn how real-time Claude Code and Codex output helps you catch drift, bad commands, and stalled agent runs before the diff grows.
Claude Code vs Codex for Local-First Development
Compare Claude Code and Codex for local-first agent workflows by task shape, context needs, review style, and Junction fit.
Track AI Coding Agent Costs Per Session
Learn why per-turn and per-session cost visibility matters when Claude Code and Codex runs become long, parallel, or automated.
Turn Local AI Agent Runs Into Pull Requests
A practical checklist for moving Claude Code or Codex work from a local agent run to a reviewable pull request.
Custom Prompt Templates for AI Coding Agents
Use custom prompt templates to make repeated Claude Code and Codex workflows clearer, safer, and easier to run from Junction.
Inspect AI Agent Runs Without Terminal Hunting
Learn what to inspect after Claude Code or Codex runs: status, output, logs, diffs, approvals, and review-ready handoffs.
MCP Server for AI Agent Orchestration
Learn how Junction's built-in MCP server lets local agents create, prompt, and inspect other agent runs without moving code.
How the Junction Daemon Keeps AI Agents Local
Understand how Junction's daemon, web app, relay, and agent sessions work together while Claude Code and Codex run locally.
Manual AI Agent Runs vs Switchboard Automation
Decide when to steer Claude Code or Codex manually and when to let Switchboard turn Linear issues into pull requests.
Why Junction Is a PWA for AI Coding Agents
See why a PWA control surface fits local AI coding agents: installability, mobile review, push notifications, and local execution.
Codex CLI vs Codex Web for Local-First Workflows
Compare Codex CLI and Codex web by execution location, local environment needs, review flow, and when Junction fits.
Use Claude Code and Codex Side by Side
Run Claude Code and Codex in separate local sessions, keep context isolated, and compare results without mixing agent work.
Write Linear Issues Ready for AI Agent Automation
Learn how to scope Linear issues so Switchboard can turn clear, bounded work into reviewable AI agent pull requests.
How to Review AI Agent Diffs from Your Phone
Learn when mobile diff review is useful, when to wait for desktop, and how to triage Claude Code or Codex changes safely.
Junction Pricing: Free vs Core vs Switchboard
Compare Junction Free, Core, and Switchboard by workflow: one daemon, unlimited chats, or Linear issue-to-pull-request automation.
Set Up Junction for Claude Code and Codex
Install the Junction daemon, pair your browser, and start controlling Claude Code or Codex from a local-first web app.
Control AI Coding Agents from Your Phone Without Moving Your Code
A practical guide to monitoring Claude Code and Codex from your phone while your code, terminals, and credentials stay on your own machine.
Git Worktree Isolation for AI Coding Agents
Keep parallel agent runs from colliding by giving each task its own worktree, branch, and review path instead of sharing one checkout.
How Switchboard Turns Linear Issues Into Pull Requests
See how Junction's Switchboard watches Linear, spawns isolated worktrees, runs agents, and opens pull requests when the work is done.
How to Approve AI Agent Actions Safely Without Slowing Down
Use permission modes, scoped approvals, and local-first execution to keep AI coding agents productive without rubber-stamping risky changes.
How to Manage Multiple AI Coding Agents Across Machines
Learn the workflow for routing AI agent work across a laptop, workstation, and VPS without losing track of approvals, diffs, or status.
How to Monitor Claude Code from Your Phone Without SSH
Keep a Claude Code session local, read live output on mobile, and decide when to approve, stop, or review the next step.
How to Monitor Codex from Your Phone Without a Cloud Sandbox
Watch Codex runs on mobile, keep execution local, and review blockers, diffs, and PR state without handing the repo to a hosted workspace.
Junction Encrypted Relay, Explained: What It Protects and Why
Understand what Junction's encrypted relay protects, what it does not, and why local-first remote access is safer than ad hoc SSH or port forwarding.
What AI Coding Agent Notifications Should Actually Tell You
Learn which agent notifications deserve attention, how to avoid alert spam, and why finish, approval, and error alerts matter most.
What an AI Coding Agent Dashboard Should Actually Do
A practical checklist for dashboards that supervise local AI agents well: live streams, approvals, diffs, notifications, and Git state.
What OpenCode Users Need From a Control Surface
OpenCode can run in the browser and attach to a terminal, but serious workflows still need remote approvals, notifications, and a broader dashboard.
Why Local-First AI Coding Agents Still Need a Control Surface
Local execution protects your repo, but you still need a browser and phone-friendly way to watch runs, approve actions, and review diffs.
Run agents where your code lives.
Install the Junction daemon, pair your machine, and keep an eye on Claude Code and Codex from any browser — no hosted sandbox required.