How-to
Keep every AI agent you run on one screen
Run several AI agents at once and by lunch you lose track of which one is waiting on you. Paste each agent’s last reply into this prompt for a board that shows which agent needs you and what to check.
Each agent ends its turn its own way: a long summary, a question buried halfway down, or a cheerful line that the tests pass. The prompt below sorts them into waiting on you, blocked, running or done, and quotes the words that decided each one.
The status board prompt
Use a fresh chat in any AI assistant. It sees only what you paste, so fill in the “about today” lines once and add one block for every agent you want on the board.
You are helping me keep track of several AI agents I run at the same time, such as coding agents, research agents or assistants working through a task. I lose time when an agent sits waiting on a question I did not see, and I lose more when I trust a "done" that was not done. Your job is to read the notes I paste from each agent and build one status board I can read without scrolling, so I know which agent needs me first and what to check before I rely on its work. You can only see what I paste here. You cannot see, contact or check on any agent yourself, so never describe an agent I did not paste or guess at what it is doing now. ABOUT TODAY (I fill this in once; any line can say "not sure") - What matters most today: [e.g. "ship the invoice rounding fix before 5pm", "the competitor research for Thursday can wait"] - Time I have right now: [e.g. "20 minutes before a meeting"] - Things I always approve myself: [e.g. "deleting data, spending money, emailing a customer, pushing to the main branch, changing anything other people use"] - My last board: [paste the board you gave me last time, or "none"] THE AGENTS (one block per agent; copy the block for each extra agent) AGENT 1 - Label: [e.g. "Claude Code, invoice-app folder", "Codex, website repo", "ChatGPT chat, competitor prices"] - The job I gave it: [e.g. "fix the rounding bug in invoice totals", or "not sure"] - Its last reply or hand-back note: [paste] AGENT 2 - Label: [ ] - The job I gave it: [ ] - Its last reply or hand-back note: [paste] SAFETY If anything I pasted contains a password, API key, token, private key, account number or a customer's personal details, say so in your first line, tell me to delete it from this chat, call it [REMOVED] from then on and never repeat it. HOW TO WORK Stage 1. Check what I gave you. If I pasted no agent notes, ask for them and wait. If a note looks cut off, or looks like the middle of a long log with no ending, write [NEED last reply] for that agent and still build the rest of the board. If a label is missing, call it Agent 1, Agent 2 and so on. If the job is blank or "not sure", work it out from the note and mark it "(my guess from the note)". Stage 2. Give each agent exactly one state, using only the words in its note: - WAITING ON ME: it asks me a question, asks for approval or offers options and stops. - BLOCKED: it hit something it cannot fix itself, such as a missing key or access, a failed install, an error it gave up on, or a usage or rate limit. Answering a question will not unblock it; I have to fix or supply something. - RUNNING: it says it is still working or will carry on. Add "as of this note" because you cannot see whether it is still going. - DONE: it says it finished. - UNCLEAR: the note does not say. Write what one line from the agent would settle it. For each state, quote the few words from the note that decided it, so I can check you. If a note says it is done but also asks an optional question (for example "want me to add tests too?"), mark it DONE and list the question as optional. If two states fit, pick the one that needs me more: WAITING ON ME, then BLOCKED, then UNCLEAR, then RUNNING, then DONE. Stage 3. For each agent, work out: - Goal: one line. - The one decision it needs from me, written as a question I can answer in a word or a choice, with the options it offered. If it is BLOCKED, write what I need to fix or supply in place of a decision. If it needs nothing from me, write "none". Never make the decision for me. - What it says it changed: files, data, messages, settings or anything it sent. If the note does not say, write "not stated". - What to check before I trust it: one concrete check tied to what it claims, such as opening the file it names, running the test or command it says passed, reading the change before it goes live, or checking two of its sources. If it says tests pass but names no command, write "ask which command it ran". If it claims nothing I can check, write "nothing to check yet". - [APPROVAL]: add this tag if the note mentions deleting, spending, sending, publishing, pushing, or anything on my "always approve myself" list, even if the agent says it is safe. Stage 4. Look across the agents. If two agents work in the same folder, file, document or account, or one agent's decision changes another agent's job, say which and why. If an agent is busy with something that clashes with what matters most today, say so. Stage 5. Build the board. - A table with the columns Agent, Goal, State (with the deciding quote), Decision needed from me, Check before trusting. Keep each cell under about 15 words. - Order the rows WAITING ON ME, then BLOCKED, then UNCLEAR, then RUNNING, then DONE. Inside each group, put the agent closest to what matters most today first. - Under the table, "Do these first": up to three actions that fit the time I have. For each WAITING ON ME agent among them, give a short reply I can paste back to that agent, with the choice left as [MY CHOICE: option A or option B] so I fill it in. Stage 6. If I pasted my last board, add a "Since last board" line for each agent: state changed, no change, or new. Flag any decision that was waiting on me last time and still is. List any agent on the last board that has no new note here as "no new note, status unknown", and never assume it finished. SELF-CHECK BEFORE YOU ANSWER - Every agent I pasted appears once, and no agent appears that I did not paste. - Every state except UNCLEAR has a quote from that agent's note, and nothing is DONE unless the agent said so. - You named no file, test, result, command or change that the note does not mention; anything missing says "not stated". - Every decision is left for me to make, and every [APPROVAL] item is tagged. - You repeated no secret I pasted.
Run it after each session
- Ask each agent for a hand-back note. Before you switch away, type into that agent’s own window: “In under 150 words: your goal, what you changed, what you are waiting on from me, and what you did not finish or check.” A short note puts the open question near the top, where a long log buries it.
- Label each note as you copy it. Name the tool and the folder or chat, such as “Claude Code, invoice-app folder”, so two sessions in the same tool stay apart on the board.
- Take out anything private. Agent output can include keys, tokens or customer details picked up while it worked, so swap them for
[REMOVED]before you paste. - Answer the top row first. Paste the ready reply the board drafts and fill in your choice.
- Check one “done” before you build on it. Run the check in its row: open the file, run the test it named or read two of its sources. A summary that says the tests pass is the agent’s word, and the board can only repeat it.
- Keep the board for next time. Save it in a note, then paste it into the “my last board” line with the new notes, and the prompt shows what changed and which questions are still waiting on you.
What a board looks like
Claude Code, invoice-app folder: Waiting on me (“Should I also recalculate past invoices, or only new ones?”). Decision: past invoices too, yes or no. Check: compare one corrected total by hand. Tagged [APPROVAL], since it changes records customers have already seen.
Codex, website repo: Blocked (“I don’t have an API key for the new email service, so I stopped”). Needs from you: the API key, added in the tool itself and never pasted in a chat. Check: nothing to check yet.
ChatGPT chat, competitor prices: Done (“Here is the final table”). Decision: none. Check: open two of its source links and confirm the prices match.
Watch them live: Oddboro
What it is: a desktop app, launched on Nate’s Substack on 13 November, where your agents work on a pixel-art office floor and where each one stands shows what it is doing.
Cost: Oddboro itself is free and open source (Apache 2.0), but it runs on Claude Code, Codex or Grok under your own account, so the AI costs whatever your plan for it already costs.
Why it fits here: version 0.0.11, out on 13 November, added a Projects panel that opens your own Claude Code and Codex sessions and marks each with a coloured dot for working, waiting for you, done or idle.
Set it up: download the file for your computer from oddboro.com/releases (Macs need Apple silicon). Setup walks you through signing in to Claude Code, your name and your office. Then click Setup under the office’s name, turn on Advanced mode and add each session under Projects with + New session, as the project sessions page describes.
On Windows: the 0.0.5 release notes said the installer was not yet signed, so Windows may warn you before it runs.
The honest bit
- The board is a snapshot: it shows each agent as of the moment you pasted, so an agent that asks a question five minutes later shows up only on your next run. Oddboro’s dots are the live view.
- Oddboro is under two weeks old: the first release came out on 2 November, Gmail connections are limited while Google’s verification finishes, and the Advanced mode warning says work sent to Claude Code or Codex runs with your permissions on your computer, outside Oddboro’s limits. Before you turn it on, check what each agent is allowed to do.
- Your check on the prompt: pick one agent the board marks done and run the check in its row. If the work does not match, tell the chat which agent, and ask it to mark that one UNCLEAR and rebuild the board.
The facts here come from Oddboro’s GitHub README and project sessions page, oddboro.com and its releases page, and the free part of Nate’s Substack launch post, read 14 November 2026.
Do it now
Before you close today’s agents, ask each one for its hand-back note, paste the notes into the prompt and answer whatever sits in the top row.
A few quick questions
Does Oddboro work with Cursor or Gemini?
Oddboro’s own site lists Claude Code, Codex and Grok as the AI an Odd can run on, and the Projects panel opens Claude Code and Codex sessions only. Cursor and Gemini are not mentioned in its FAQ. For any other agent, the prompt works with whatever it writes at the end of a turn.
How many agents can one board handle?
As many as fit in the chat, but each note adds length, so ask each agent for the short hand-back note and leave out long logs. If the board gets too long to scan, split it by project and run the prompt once per project.