The Short Answer
CanvasTTY is an open-source Electron desktop that arranges live terminals and AI-agent CLI sessions on an infinite canvas, zooming out to readable summaries rather than unreadable thumbnails. It is a developer tool, but it answers a problem every organization running concurrent agents will meet: a linear chat transcript cannot represent parallel work. On ibl.ai you own all the code and the data, including the interface layer, so the supervision model is yours to change.
The chat window is not wrong. It was designed for one assistant answering one question, and it is excellent at that. It simply does not extend to supervising many agents at once.
Why does a chat window stop working once agents run in parallel?
A chat window stops working because a transcript is linear and sequential while concurrent agent work is neither.
With one assistant, the conversation is the task, and reading top to bottom reconstructs everything that happened. With five agents running against different parts of a workflow, the questions you need answered are structural rather than narrative:
- Which agent is blocked, and on what
- Which one is waiting on a human approval
- Which one is consuming budget faster than expected
- Which one has quietly drifted into doing the wrong thing
None of those is answerable at a glance from a scrolling log. You get them by reading, which is precisely the cost that automating the work was supposed to remove. Tabs do not fix it either — a tab hides state by design, so a failure in tab four is invisible until you click it.
This is the interface bottleneck that arrives immediately after an organization gets its first agents into production, and it is under-discussed relative to how quickly it bites.
What does CanvasTTY do differently?
CanvasTTY treats sessions as locations rather than as entries in a list, which is what makes many of them viewable at once.
It is an Electron spatial desktop for real local PTYs and AI-agent CLI sessions. You launch a shell or an agent in a project directory, position and resize its live terminal on an infinite canvas, and navigate by moving around rather than by cycling through tabs.
A fixed Home zone holds sessions, limits, and launch shortcuts. Its tagline states the model directly: your terminals are places, not tabs.
The design decision worth extracting is what happens at distance. Zoomed out, live terminals do not shrink into illegible noise — they render as readable summaries. Distance maps onto abstraction level.
That is a supervision interface rather than a display trick. It lets one person hold ten concurrent workstreams in peripheral view and descend into one without losing awareness of the other nine, which is exactly the operation a linear transcript makes impossible.
CanvasTTY is at version 1.0.2, ships for Linux, Windows, and Apple Silicon macOS, and credits DriftWM as its inspiration.
What CanvasTTY is not
CanvasTTY is a developer tool for terminals and agent command-line sessions, and it is worth saying so plainly because it is easy to over-read.
It is not an enterprise agent platform, and it is not an education product. It has no classroom features, no institutional identity integration, and no governance layer. Describing it as an ed-tech breakthrough misrepresents what its authors built.
What it is: an early, concrete demonstration of a spatial supervision model for concurrent agents, in a domain — developer tooling — where people run many agents in parallel today and therefore hit the problem first.
That makes it a useful preview rather than a product recommendation.
Why is the interface layer a strategic question rather than a cosmetic one?
The interface layer is strategic because it determines how many agents one person can actually supervise, and supervision capacity sets the ceiling on deployment.
An organization does not get value from agents in proportion to how many it runs. It gets value in proportion to how many it can run safely — with a human able to catch the one that went wrong before the consequence lands.
If your interface lets one operator credibly oversee three concurrent agents, three is your ceiling regardless of what the platform can technically launch.
Improving that ratio is an interface problem, not a model problem. And it is the reason the supervision surface deserves the same scrutiny usually reserved for model selection.
There is an ownership consequence. You can only reshape the interface if you own the client. A per-seat SaaS assistant ships the vendor's chat window, and the supervision model available to you is whatever they built for the median customer.
When your operators need a different view — grouped by risk, by approval queue, by cost, by business process — you can file a feature request.
The per-seat cost shape makes it worse in the same direction: at roughly $30–60 per user per month you are paying by headcount for an interface whose limits are setting your agent ceiling.
On ibl.ai you own all the code and the data, run it model-agnostic across any LLM, and pay with no per-seat pricing — so you can deploy anywhere, and change the interface your operators work in. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.
For the organizational prerequisite that determines whether agents reach production at all, see why 99% plan to deploy AI agents and 9% have.
Frequently asked questions
Is CanvasTTY an enterprise product?
No. It is an open-source developer tool at version 1.0.2 for running local terminals and agent CLI sessions on a spatial canvas. It is worth studying as an interface pattern rather than deploying as institutional infrastructure.
Does a spatial interface actually improve agent supervision?
The mechanism is plausible and demonstrated in a narrow domain: rendering distance as abstraction lets an operator keep many concurrent sessions in view.
Whether it improves outcomes at organizational scale has not been established, and we are not aware of published evidence either way.
What should an organization take from this today?
Measure your supervision ratio — how many concurrent agents one operator can credibly oversee in your current interface. That number, not model capability, is likely what caps your deployment.
The bottom line
The next meaningful improvement in agentic AI for most organizations will not come from a better model. It will come from being able to supervise more agents at once without losing the thread.
CanvasTTY is a small, specific answer to that in a developer context. The general version — a supervision surface shaped around how your work actually runs — is only available to organizations that own the client they are looking at.