Agents
An agent block is the interface for watching and steering an AI agent that does work on someone's behalf: the steps it takes, the apps it touches, and the moments it stops to ask. Orvain's Agents family runs to eighteen, from a single run on a rail of steps with its approval in the row, through a plan read before it starts, checkpoints to branch from and parallel attempts to choose between, to fleets, lanes and queues for many agents at once, a trace and a cost waterfall for where the time and money went, and an overnight recap and an inbox for everything that waits on a person.
The hard part is not showing that an agent is busy; it is showing what it is about to do while there is still time to say no. Every block here puts the decision beside the thing it decides: the command it will run under the step that runs it, the diff a signer is signing, the rule that set a gate. A run that only reports what already happened is a log, and nobody can steer a log.
18 blocks
Using Agents
Reach for it when
- An agent that acts in other people's apps, where some of its actions need a yes before they happen.
- Several agents or several attempts running at once, where someone has to see which is stuck and which is done.
- Runs long enough that people leave and come back, and need to catch up on what happened and what waits on them.
- Spend or time that has to be explained afterwards, step by step, against a budget.
Reach for something else when
- A single question and a single answer. An assistant thread from the AI Chat family is the right surface, with its research shown inline.
- A background job with no decisions in it, such as a nightly export. A progress bar and a notification say all there is to say.
- A fixed automation people design ahead of time. Build it on the Workflow canvas, where the steps are drawn before they run rather than chosen as it goes.
Where it shows up
Release agent
An agent that tags, builds and publishes a release. The monitor shows each step sized by how long it took, and the publish step waits in its row with the exact command until someone allows it, so nothing reaches customers on the agent's judgement alone.
Support fleet
Four agents working the refund queue overnight. The fleet shows which lead started which sub-agent, the queue shows jobs waiting on rate limits, and in the morning the recap lists what was done and the two refunds that still need a person.
Questions
How should an AI agent ask for approval?
In the row of the step that needs it, with the exact action it will take: the command, the amount, the message it will send. A separate dialog that says “The agent needs approval” forces the person to reconstruct what they are approving. Show allow and decline together, say what each one does next, and keep the decision visible afterwards so the run explains itself later.
How do you show many agents running at once?
Group them by what they share. Agents working the same goal sit under the lead that started them; agents working through the same pool of capacity sit in a queue with their workers; agents compared over the same stretch of time sit in lanes on one time axis. Each grouping answers a different question, so pick the one people are actually asking, which is usually which agent is stuck and why.
Premium components,
free in your inbox.
Join the Orvain Dispatch and get production-ready components, free to keep: React and Vue code plus the Figma frame, yours to ship.
Orvain MCPAgents
An agent block is the interface for watching and steering an AI agent that does work on someone's behalf: the steps it takes, the apps it touches, and the moments it stops to ask. Orvain's Agents family runs to eighteen, from a single run on a rail of steps with its approval in the row, through a plan read before it starts, checkpoints to branch from and parallel attempts to choose between, to fleets, lanes and queues for many agents at once, a trace and a cost waterfall for where the time and money went, and an overnight recap and an inbox for everything that waits on a person.
The hard part is not showing that an agent is busy; it is showing what it is about to do while there is still time to say no. Every block here puts the decision beside the thing it decides: the command it will run under the step that runs it, the diff a signer is signing, the rule that set a gate. A run that only reports what already happened is a log, and nobody can steer a log.
18 blocks
Using Agents
Reach for it when
- An agent that acts in other people's apps, where some of its actions need a yes before they happen.
- Several agents or several attempts running at once, where someone has to see which is stuck and which is done.
- Runs long enough that people leave and come back, and need to catch up on what happened and what waits on them.
- Spend or time that has to be explained afterwards, step by step, against a budget.
Reach for something else when
- A single question and a single answer. An assistant thread from the AI Chat family is the right surface, with its research shown inline.
- A background job with no decisions in it, such as a nightly export. A progress bar and a notification say all there is to say.
- A fixed automation people design ahead of time. Build it on the Workflow canvas, where the steps are drawn before they run rather than chosen as it goes.
Where it shows up
Release agent
An agent that tags, builds and publishes a release. The monitor shows each step sized by how long it took, and the publish step waits in its row with the exact command until someone allows it, so nothing reaches customers on the agent's judgement alone.
Support fleet
Four agents working the refund queue overnight. The fleet shows which lead started which sub-agent, the queue shows jobs waiting on rate limits, and in the morning the recap lists what was done and the two refunds that still need a person.
Questions
How should an AI agent ask for approval?
In the row of the step that needs it, with the exact action it will take: the command, the amount, the message it will send. A separate dialog that says “The agent needs approval” forces the person to reconstruct what they are approving. Show allow and decline together, say what each one does next, and keep the decision visible afterwards so the run explains itself later.
How do you show many agents running at once?
Group them by what they share. Agents working the same goal sit under the lead that started them; agents working through the same pool of capacity sit in a queue with their workers; agents compared over the same stretch of time sit in lanes on one time axis. Each grouping answers a different question, so pick the one people are actually asking, which is usually which agent is stuck and why.
Related blocks
All blocksPremium components, free in your inbox.
Join the Orvain Dispatch and get production-ready components, free to keep: React and Vue code plus the Figma frame, yours to ship.
