Pulse - We automated work inside legacy software that has no API, from booking appointments to monitoring records, without anyone needing to operate the software manually.

24/7Autonomous operation

1,000+Production tasks

150+ hrsManual work eliminated

3+Live workflows

The problem

Some of the most important software in a business is also the hardest to automate.

Desktop applications, remote environments, legacy healthcare systems, and internal tools often expose no usable API, webhook, or integration layer. The work still happens through a screen, which means a person has to sit there and do it. Pulse was built for that gap. It gives AI agents a persistent computer environment where they can operate software directly, understand its current state, execute multi-step workflows, recover from failures, and leave a complete audit trail behind. Today, Pulse runs continuously across multiple production workflows. In dental operations, it handles both outbound workflows for existing patients and inbound workflows for new patients, following different paths depending on the workflow. In other deployments, it handles monitoring and analysis inside legacy systems that would otherwise require substantial manual review.

How it runs · 08 steps

Appointment / Task ArrivesA new task enters Pulse from the surrounding workflow or application.

trigger

Redis Job QueueFastAPI validates the payload and pushes it to Redis, where available agent instances wait for work.

process

Agent Claims JobOne instance atomically claims the task, preventing multiple agents from processing the same job.

agent

Desktop Session OpensThe agent connects to the remote environment containing the legacy software.

process

Preflight CheckClaude reads the current screen and determines whether the environment is ready or additional navigation is required.

Computer-Use ExecutionClaude operates the software directly, navigating the interface and completing the required workflow based on the current screen state.

agent

Session RecordedThe complete desktop session is captured and stored alongside the task for review and traceability.

storage

Confirm & SyncThe result is written back to the surrounding system and the appropriate downstream notification or action is triggered.

output

Engineering detail · 04

Computer use instead of integrationEXACT and similar legacy systems expose no usable integration surface, so Pulse gives Claude the same interface available to a human operator — the desktop itself. The agent reads screenshots, understands the current UI state, and chooses the next action rather than relying on a predefined sequence of screen coordinates.

Parallel by designIndependent agent environments consume jobs from a shared Redis queue. Each has its own desktop and RDP session, while atomic job claims prevent duplicate processing. If one agent becomes unavailable, the others continue working.

Observable by defaultEvery desktop session is recorded from start to finish and attached to its task. Model usage and execution metadata are tracked alongside it, providing visibility into what happened, what the agent saw, how long it took, and what it cost.

Failure is part of the architectureTasks can retry automatically. Stale jobs are recovered on startup, orphaned work can be detected and requeued, and running agents can be cancelled safely without leaving the queue in an inconsistent state. The goal is not an agent that works in a demo. It is a system that can keep operating when nobody is watching.

Result

Pulse has moved beyond a single automation into infrastructure that runs continuously across multiple production workflows. In its first live dental campaign, it booked 7 of 7 appointments with zero failures at $14.05 total model spend. Today, the same underlying architecture supports inbound and outbound booking flows as well as monitoring and analysis workflows inside legacy software. Work that previously required people to repeatedly navigate desktop systems, transfer information, and review operational data can now run autonomously, with every execution leaving behind its status, session recording, and usage history. The system doesn’t demonstrate computer use. It’s part of the operation.

”

“The booking agent handles what used to take our coordinator 30 minutes every morning. Seven bookings on the first live run, zero failures. It just runs.”

Operations Lead · UK Dental Group · Private

NextAtlas →Building and deploying a new AI agent required a technical person. Atlas lets anyone on the team turn a client brief into a working agent in under five minutes.

Bring me a problem like this →