A modern runtime for agents
Define agents, give them skills, and watch them do real work in isolated sandboxes.
Give your agents a place to actually work.
A chat box is not a runtime. Agents need triggers, a sandbox to run in, a way to be tested, a way to be reused, and a gate before their work reaches production. JuhJuh gives them all of it and keeps a human on the approval.
An agent is not a chat session. It is a worker with a job.
Triggers
Fire on an event, on a schedule, or on a button.
Playground
Test an agent in chat before you put it on real work.
Pipelines
Chain agents so one's output is the next one's input.
Batch runs
Point an agent at many items and let it work through them.
Marketplace
Install ready-made agents instead of building from scratch.
An agent is the worker. Its Skills are the toolkit it carries to every Task.
Skills are reusable capabilities, things like review-fix, generate-prd, and run-tests. Build your own in the editor, publish them for your workspace, or pull from the template library.
A bank's worth of agents, each with a real job.
Meridian runs a roster of agents themed to its own services. Each one carries the Skills for its job and hands its work back for review.
Fraud Triage
Scores incoming transactions, flags the suspicious ones, and opens a Task with the evidence attached.
KYC Reviewer
Runs OCR on identity documents, checks the fields, and routes the edge cases to a human officer.
Dispute Resolver
Gathers the chargeback paperwork, drafts the response, and keeps the case moving to its deadline.
Support Copilot
Drafts replies to customer tickets from the knowledge base, then waits for an agent to send.
Transaction Categorizer
Labels spend across accounts so statements and Smart Savings Pots stay tidy and searchable.
Compliance Scanner
Reads each change for regulatory risk and raises a Task before anything reaches a customer.
The agents move fast. Your team keeps the judgment.
Agents do the work in isolated sandboxes and hand back a result: a pull request for code, a finished artifact for everything else. They do not merge and they do not publish on their own. Every result lands in front of a human for review, and code changes brew through the Cauldron first.
- Work runs in isolated sandboxes
- A pull request for code, a finished artifact for everything else
- No autonomous merges, no autonomous publishing
- Every result reviewed by a human
- Code changes brew through the Cauldron first
Watch one agent work a Task, line by line.
Because each Task runs in its own isolated workspace, many agents can work at the same time without colliding. A whole roster can be in flight while your engineers review the results. Here Fraud Triage picks up a flagged transaction, scores it, gathers the evidence, and hands back a Task for a human to decide.
- Its own sandbox, no collision with other runs.
- The result lands as a Task, with the evidence attached.
- A human makes the call before anything reaches a customer.
Give your agents a place to actually work.
See how Cauldron, the CI/CD for agents, gates every change before it ships.
