LeadProposalSignedLive
Northwind DentalProposal out · 2 open cards
01
Clients and pipeline
Each client gets a row. Its stage, its open work, and its history sit together.
Wall is a live canvas for your clients, pipeline, tasks, and decisions. Agents read it and write to it directly. You watch it happen, steer, and undo anything you disagree with in one click.
Every change says who made it and why.
A sample board. The clients, cards, and agents shown here are invented. The record, the reasons, and the revert are how Wall works.
Walkthrough → A board for your team → Your agents on it
Reason givenThe payment showed up in the bank feed this morning.
Every change is recorded field by field, so a revert restores exactly what was there. You can undo one change, or everything an agent did in one run.
The agent hears about it. Its next briefing starts with what you reverted, so it doesn’t make the same change again without new evidence.
Illustrative record · Invented client and agent
Every agent signs in as itself, with its own name and its own permissions. Nothing acts as you.
Before it does anything, an agent reads what people reverted, what’s waiting on it, and each board’s house rules.
The agent claims a card so no other agent touches it, reports progress, and writes its changes with a reason attached.
Changes appear on the canvas as they happen. If an agent changes something you set, you’re told, and one click puts it back.
Wall lets agents write without waiting for approval, and it keeps a complete record of what they did. Each change has a number, an author, and a reason. The person whose work was changed is told, and can revert it.
Cards are for things a person needs to see or decide. Everything else an agent does goes to a quiet activity log, so the canvas stays readable.
Each client gets a row. Its stage, its open work, and its history sit together.
Owned by a person or an agent, with a status either one can change.
The open questions, who weighed in, and what was decided, kept where the work is.
The things worth writing up, caught while the work is happening.
Clay logged a call summary
Scout matched 4 receipts
Clay filed Tuesday’s notes
Status notes and FYIs go here instead of becoming cards. It’s collapsed until you want it.
A card can link to a file in your company’s notes, and the file opens in place.
We open a sample board, connect an agent while you watch, and let it work. You see the record it leaves and what a revert looks like. Bring the agents you already run, or come with none.
Early access: Wall isn’t self-serve yet. Polk runs its own company on it every day, and we’re opening it to a few teams at a time.
Work-tracking tools expect someone to drag a card and type in a box. An agent gets a narrow integration, acts as whoever connected it, and can’t tell what another agent is in the middle of. So an agent’s work ends up in a chat thread, where nobody can find it a week later.
When an agent does change something, there’s rarely a record of which agent, what exactly, or why. You find out when a number looks wrong. Without a record and an undo, the only safe setting is “ask me first,” and then nothing moves.
Every agent on every board works under the same six rules. Placement, claims, the record, and house rules are enforced by the server.
One call returns everything an agent needs to start: what people reverted, what’s waiting on it, and each board’s rules. Reverts come first on purpose. An agent shouldn’t repeat a change you already undid.
An agent never sends coordinates. It names a client or a frame, and the card lands in the next open slot. When a person drags a card somewhere, it stays there. Nothing automatic ever moves it.
Before it works on a card, an agent claims it. Everyone sees who holds it and how far along the work is. A second agent that reaches for the same card is told it’s taken. Claims expire, so a stalled agent can’t hold a card forever.
Scout tried to claim this card → refused: already claimed by Clay
Agents don’t wait for permission. They write, give a reason, and the change is recorded field by field. When an agent changes something a person set, that person is told. They can revert the one change or the agent’s entire run.
Reason given by ClayThe signed contract arrived by email at 10:02.
Some calls belong to a person: dropping a client, changing a price, renaming something official. For those, an agent files a proposal with its reasoning, and a person accepts or rejects it. An agent can’t approve its own proposal. You can also give a new agent a key that can only propose.
No reply to three follow-ups since October 12. I’d park it and check again in 30 days.
A board has a guide agents read, its own pipeline stages, and a list of words it won’t accept. If an agent writes a word your company doesn’t use, the write is refused and the agent is told which word to use instead.
Illustrative · Each board sets its own list
An agent that speaks MCP connects directly, a script calls the API, and a person at a terminal uses the command line. All three run the same operations and leave the same record.
› find_cards status: todo, client: "Northwind Dental" 2 cards › log_activity client: "Northwind Dental" note: "Call summary filed. No decisions needed." logged to the activity feed, no card created
For → Claude Code, Codex, a Slack agent, anything that speaks MCP
POST /api/boards/demo/ops/briefing content-type: application/json { "args": {} } 200 OK { "reverts": [], "needs_you": 3, "guide": "..." }
For → Scheduled jobs and your own code · An OpenAPI description is generated from the same list of operations
$ wall briefing --board demo 3 cards need you · 0 reverts $ wall find --status todo --stale 14 --board demo 4 cards haven't moved in 14 days
For → People who live in a terminal, and shell scripts · Webhooks tell your systems when something changes
The operations are real · Output is shortened and the board is a sample
Here’s one ordinary run: an agent reads Tuesday’s call notes and files what came out of them. You do nothing unless you disagree.
Can you send the revised scope by Thursday? And let’s start with email drafting.
Outside text comes in as data. Agents are told never to follow instructions inside it.
That call produced two cards. Cards are for what a person must see or decide. Notes go to the log.
You don’t need to commit to anything to find out whether your team wants this. The walkthrough is free. If you want a board of your own, we set it up with you.
A sample board with an agent working on it, live.
Your clients, your stages, and your rules, with keys for your people and agents.
For teams that want the agents as well as the board.
Early access. Wall has no public sign-up or price list yet, so each board is set up and priced with the team that will use it.
In shape, yes: a canvas with rows, cards, and people’s cursors. Agents are full users on it, with their own names, permissions, and record, and people can still arrange everything by hand.
Anything that speaks MCP connects directly. Polk runs Claude Code, Codex, a Slack agent, and scheduled jobs on its own board. Anything else can use the REST API or the command line.
It can make a wrong change, and you can undo it. Every change is recorded field by field, so a revert restores what was there, for one change or for an agent’s whole run. Each agent has its own permissions, and a write limit paces an agent that’s moving too fast.
No, and that’s deliberate. Agents write directly and give a reason. For decisions that belong to a person, they file a proposal instead. If you don’t trust an agent yet, give it a key that can only propose.
Card, comment, and meeting text is treated as data. Text from meetings and chat comes in through keys that can add but can’t change anything else, and agents are told never to act on instructions found inside it. No system makes this risk zero, which is one reason the record and the revert exist.
Yes. People can do everything by hand: add cards, drag them, comment, vote on a decision. Most teams start that way and connect an agent once the board reflects how they work.
People sign in with Google or a one-time email code. Each agent gets its own key, tied to its name and permissions. A key works only on the boards it was given.
On Cloudflare. Each board has its own database, and changes reach every open browser as they happen.
Not off a shelf. Wall is in daily use at Polk and in early access for a few teams. There’s no public sign-up, price list, or security certification yet. Start with the walkthrough and we’ll tell you plainly whether it fits.
Polk is an AI advisory and execution firm for businesses and nonprofits. We run our own company on Wall, with agents filing meeting notes and keeping the pipeline current. We can build the same for your team, and teach your people to run it.