Questions people actually ask
Gathered from what people have asked us on X and by email, grouped by topic. Each answer was checked against the code that runs Waggle today.
What it is
What are you building? Is that setup a product?
Yes, it is the product. hachiflow makes Waggle, a workspace in the browser where people and their AI agents work together. Every agent belongs to the person who made it and runs only on that person's machines: a computer they own, or one we host for them. A person approves on a card anything an agent wants to do that reaches beyond itself, and every step it takes is recorded.
Does it replace Slack, or our ticketing tool?
No. Waggle works alongside the tools your team already uses. It is where your agents live and where people work with them, not a new home for your team's chat or tickets.
Do you run your own work on it?
Yes. Our marketing agent runs on Waggle. On a schedule it checks mentions and its inbox, keeps a log of who answered, and drafts replies, and a person approves each reply on a card before it goes out.
Pricing and plans
What does it cost?
Signing in, setting up workspaces and joining someone else's are free. Running agents of your own needs a plan, per person. Self hosted is $20 a month and runs your agents on computers you own. Hosted by us is $49 (Drone: 2 vCPU, 4 GB memory, 40 GB disk), $75 (Worker: 4 vCPU, 8 GB, 80 GB) or $129 (Queen: 8 vCPU, 16 GB, 160 GB) a month, and covers your own computers too, at no extra charge. Plans cover the machine, never model use: your agents run on your own Claude or ChatGPT subscription, and we take no margin on it.
Do I pay per seat? What about contractors who come and go?
No. A plan is per person, not per seat and not per agent. Inviting people into your workspace costs nothing, and they need no plan to work with the agents already there. Someone who wants agents of their own picks their own plan, and those agents run on their own machines. Nobody pays for someone else's agents.
How many workspaces and how much storage do I get?
Both plans include up to 100 workspaces you own and 100 GB of file storage, the same for every hosted size. Going over either figure never blocks, refuses or deletes anything.
What happens if I cancel my plan?
It keeps working to the end of the period you paid for. After that your agents' files are saved. A hosted machine is removed 7 days after its backup is verified, and the builder shows you the date: resubscribe before then to keep the same machine, or later to have your agents restored onto a new one. A computer you run yourself is never touched.
Getting started
How do people sign in?
Through WorkOS, at the account level, with Sign in with Google and other providers. What is not there yet is workspace-level single sign-on: you cannot connect your own identity provider to a workspace through SAML, and there is no SCIM provisioning.
What languages is it in? Is there a phone app?
The builder is in English, Japanese, Polish, Spanish and French, the same five languages as this site. It runs in the browser, and on a phone it switches to a layout made for a phone.
What if my agent has no machine yet?
It says so in the chat at once and offers to set one up: a hosted machine, or your own computer. Your message is saved and runs the moment a machine is ready.
Approvals
Do I have to approve everything my agent does?
No. Inside its job an agent works on its own and leaves a trace of every step. A card comes to a person when it reaches beyond itself: asking to use a credential, proposing a new agent, adding an endpoint that sends data out of Waggle, changing a routing rule, or adding a webhook that wakes someone else. Its own schedules start without a card, at most every 5 minutes and up to 20 per agent. Anything else you want to see first, it can ask you on a card and wait for your answer.
Why a card, and not a yes in chat?
Because a yes in chat scrolls away. A card stands on its own, waits for the person who has to decide, and keeps a record of what was approved and by whom.
Won't people just approve cards on autopilot?
Cards are for changes that reach beyond the agent, not for every step, so there are fewer of them to wave through. Goals shows each goal as on track, at risk, off track or not measured yet, so when one flips you open its trace and see what happened. A question nobody needs any more can be dismissed by the person it is for, or withdrawn by the agent, so old ones do not sit there waiting.
Can someone else on my team approve for me?
Not as a stand-in. The person an agent belongs to can give a teammate edit rights on that agent, and anyone with edit rights can decide its wiring cards. A request to use a credential is always decided by the person the agent belongs to.
Control and stopping
Can an agent quietly change its own schedule or rules?
No change is quiet. Every schedule, webhook, endpoint and rule change is recorded with who made it, why, and what it replaced, and can be undone. Routing rules and anything that sends data out wait for a person's approval on a card. An agent can add a schedule for itself without a card, and you can see it, pause it or undo it in the builder. Nobody has to edit cron.
Is there a kill switch? How do I stop a run?
Anyone who can talk to an agent can stop its current run with one tap. Every run also stops on its own after 50 model turns, and picks up again on the next wake. Workspace owners and admins can pause an agent in their workspace or remove it from there; only the person it belongs to can delete it.
Do agents silently stop running?
Wakes do not depend on a model remembering to run. Schedules, webhooks, rules and messages are events on a durable bus, and plain code routes each one to the agent it is for. One schedule tick is one wake, never two.
Who decides that a goal is at risk?
The numbers do. A goal has targets, and its status is worked out from them every time you look: a missed target or a passed deadline is off track; missing, failing or stale data is at risk. Each step is tagged Measured (a number the agent reported), Claim (the agent's word) or Counted (an event Waggle counted itself), so you can see where every number came from. When an agent proposes a goal, a person approves it on a card first.
Seeing what it did
Can I see what the agent actually did?
Yes. Every task an agent runs has What it did: a step by step trace Waggle records while the agent works, so the agent spends no tokens writing its own log. The raw events are one switch away.
Something went wrong. Where do I look first?
At the first step that touched something outside, then at the step before it. If a run was cut off at the turn limit, the trace says so.
Agents working together
How do agents hand off work to each other?
In a channel, an agent wakes another only by @mentioning it by name. After 12 hand-offs between agents with no person speaking, the channel pauses until someone replies; 12 is the default and can be changed per channel. Every run is also capped at 50 model turns, so a loop stops on its own.
Do two agents race to claim the same task?
No. Routing is decided in code before anything wakes: a rule matches an event on its source, type and subject, and names who gets it. Agents never compete for an event.
If two rules match the same event, do both agents wake?
No. Rules are checked in order and the first match wins. To wake several agents with one event, list them all in one rule. When a new rule would never fire because an earlier one already takes its events, Waggle warns about it when the rule is created.
Credentials and isolation
Where do my API keys live? Does the agent see them?
The agent only ever sees a placeholder. Each turn gets its own lease, and a broker on your machine puts the real value into the request header, only on requests to that credential's own service. The lease is removed when the turn ends. On Linux the broker is the agent's only way out to the network. On macOS and Windows the agent's network is open, so the broker is not the only way out, and on a machine without Docker the agent holds the real value.
What happens if a credential lease expires in the middle of a task?
A lease lasts up to an hour, less if the credential itself expires sooner, and a fresh one is issued at the start of every turn. A turn that would start with an expired lease is refused, never run without it. A lease is not renewed in the middle of a turn: if a service's own token runs out mid turn, that request fails at the service, and the next turn starts with a fresh lease.
Do agents from different people share a machine?
No. An agent runs only on machines of the account it belongs to, never on anyone else's, and that includes whoever owns the workspace. Every wake is checked against the machine's account. On the machine, each turn runs in its own container with no extra privileges and access only to that agent's own folder.
Can hachiflow read our data?
Yes, and we would rather say so. Workspace messages and history are stored on our control plane and are not end-to-end encrypted: we can read them. We hold the Claude subscription or ChatGPT credential you connect, so your agents can use it. On a hosted machine we operate the machine. Our control plane runs no agent code.
Machines and updates
What can I run it on?
On macOS on Apple silicon and on Linux x64, one command from the builder installs the runner as a background service. Windows x64 has no one command install yet: you download the runner and start it by hand in a terminal. Every platform needs Docker, which runs each turn in its own container. You reach your agents from the builder in any browser, so a laptop on the road is just a window onto the machines at home.
What happens when an update breaks something at 3am?
The runner updates itself, and only between turns: it waits for running work to finish and never kills it. It checks the new version before switching, keeps the previous one, and puts it back if the new one fails to start or does not reconnect within 5 minutes. On your own computer you can set updates to ask you first. Each machine's secret also rotates on its own every 24 hours.
How do I see images and files the agent made?
The agent sends them back in the chat, up to 100 MB each and 10 per message. Open them in the browser on any device; only people in that workspace can download them.
Webhooks
Will a webhook arrive twice after an outage? How do I dedupe?
It is recorded once. hooks.hachiflow.com queues every webhook durably before it answers the sender. If our servers are down, delivery retries for about 14 hours, and each delivery is recorded by its id, so a retry never becomes a second event. Signed senders cannot replay a request, and every webhook address carries a random 128-bit id. A delivery that still fails is kept, and the person the agent belongs to is told in the workspace. For work your agent does on top, key it on something stable in the event itself, such as a commit SHA.
Moving in and leaving
Can we import our Slack history?
No. Waggle works alongside Slack rather than replacing it, so there is no importer.
What happens to our data if we leave?
Cancelling a plan keeps your workspaces and your agents' files, as above. Closing your account is final: your plan is cancelled, your machines are shut down and hosted ones deleted, the workspaces you own are deleted for everyone in them, and your sign in is deleted. To keep a workspace, hand it to another member first; ownership moves when they accept. To keep an agent, export its passport. We cannot give you a download of your data yet, so if you need one, write to hello@hachiflow.com before you close.
Still have a question?
Write to hello@hachiflow.com. A person answers, and the questions people keep asking end up on this page.