The Pushary blog
Guides
How to run AI agents safely, keep a person in the loop, and stay in control.
Claude Code approval fatigue: what the numbers say and what to change
People click yes on almost every Claude Code prompt, and they catch less the longer a session runs. Fewer, better questions fix that. More prompts do not.
Codex stuck on Awaiting approval: what it means and how to fix it
Most stuck chats are a request you have not seen. Some are a bug where the request cannot be answered. How to tell them apart and get Codex moving again.
How long AI agents wait for people: the 2026 data
Agents work for minutes, then stop for a person. Much of the wait is a question nobody has seen yet. Every number here is public and dated.
How to roll out Codex managed settings to a team
A Codex requirements.toml example, where to deploy it, and four checks to run with a teammate before rollout.
Tools that let a team approve an AI coding agent's risky steps, compared
Claude Code, Codex and Cursor only ask the person at the keyboard. These tools can ask a lead instead. Which agents each one covers, where approvers answer, what it records and what five people pay.
Can you route a Claude Code permission prompt to your team lead before it runs?
Not with Claude Code alone. A permission prompt goes to whoever is at the terminal. Here is what a hook can do, what managed settings can't, and how a routing rule sends the step to a lead.
An AI agent broke something. How to trace what ran and whose yes let it
Git names the machine, the session names the command, and only a recorded approval names who said yes. Where to look, in order, and what each place can and can't tell you.
Claude Code managed settings for a team: a working file, and what it can't do
The one Claude Code file your team can't override: where it goes, a version you can copy, how its rules are read, and what it still can't do for a team.
Who approved it? What Claude Code, Codex, Cursor and Copilot team plans record
Each tool lets the person running it say yes. None of them, on any team plan, documents asking a second person first. Here is what each one does record.
How to add human-in-the-loop to a Vercel AI SDK agent
The AI SDK gives your agent tools. One of those tools can be a real person. Here is the whole integration, from install to fail-closed decision.
LangGraph human-in-the-loop: interrupt, approve, and resume from a phone
LangGraph pauses the graph and persists state. It does not notify anyone. Here are the two patterns that finish the job, with the footguns written down.
How to approve Claude Cowork tasks from your phone
Paste one connector link, add standing instructions, and Cowork's questions arrive as push notifications you answer from the lock screen.
An audit trail for Claude Cowork: what did your AI coworker decide while you were out?
Delegation without a record is hope. Keep every Cowork question and answer in an immutable, exportable trail, next to every other agent you run.
Claude Cowork stuck waiting for permission? Stop babysitting it
You hand Cowork a task and walk into a meeting. It hits a question at step two and waits. Here is how to get those questions on your phone instead.
What an AI agent audit log should capture for teams and compliance
The fields a coding-agent audit record needs to be worth keeping, and the honest line on what GDPR-aligned and self-assessed actually means.
Who is accountable when an AI agent makes a mistake?
An agent has no accountability of its own. The human who ran it owns the outcome, which is why a record of who approved what matters.
How to run multiple AI agents at once without losing track
The workflow and the board for running concurrent agent sessions without losing track of which one needs you.
AI agent control glossary: HITL, MCP, permission gates, kill switch, audit trail
Each term gets a one-line definition first, then context and a short example. Built so you can lift any single entry.
Vibe coding safety: letting AI run code without babysitting it
The real risk in vibe coding is the commands an agent runs, not just the code it writes. Scanners miss that. Human approval on actions closes it.
How to run an AI agent overnight and stay in control from your phone
Let the agent work while you sleep, but keep the gates. Risky steps route to your phone for a yes or no, and a kill switch ends a bad run.
How to stop an AI agent from running up your token bill
A max-budget flag warns you. A budget that kills the session and an approval gate stop the bill from growing while you are away.
What --dangerously-skip-permissions does, and a safer way to run unattended
Let routine work continue. Choose which commands need your answer, and know what an approval hook cannot protect.
What permissions should an AI agent have? A starter least-privilege policy
Reads auto-approve, writes and pushes and spends ask, secrets and force-pushes deny. A concrete starter policy you can copy.
The four levels of AI agent oversight: notify, approve, policy, audit
Oversight comes in four levels. Notify is parity now. Approve catches the moment. Policy decides without you, including on the amount. Audit proves what happened.
Is it safe to let an AI agent run unattended?
It depends on what you put around it. Unattended is safe when reads run free, risky actions gate to your phone, and a kill switch and audit trail are in place.
Allowlist vs denylist for AI agent permissions, and why prefix denylists leak
A denylist tries to name every dangerous command and always misses one. Pair a read-only safe floor with an allowlist instead.
What is MCP, and how agents use it to notify you and ask for approval
The Model Context Protocol lets agents call external tools. Two of them can ping you and ask for a yes or no before the agent acts.
Human in the loop for AI agents, explained
The agent works on its own and stops only for the few decisions a person should see. Here is what that means and where the checkpoints sit.
Can an AI agent delete your files or drop your database?
Yes, an agent with shell or database access can delete your files, and it has happened in production. Here is how to gate it.
Let your agents work. Stay in control.
Get the questions that need you on your phone or Mac. Approve, reject, or answer so your agent knows what to do next.