The work agents are good at, and the reason it stays manual
Updating 1,200 customers by a rule, filling in missing fields, moving records through a process, sorting out what came in overnight: this is work an agent can do. Most teams keep it manual anyway, because a mistake in live data is expensive and hard to find.
Say an agent updates those 1,200 records and three get the wrong plan. Nothing errors. Somebody notices weeks later, when a customer is billed for a plan they never chose. The usual advice, keeping agents read-only or approving each write, means either no help or 1,200 clicks.
How it works in CodeKasten
- Keep the work in a workspace. Build the record types and views your team needs, or install an app from the Marketplace.
- Give the agent its own branch. It works on a full copy, and
mainstays protected. It can do the whole job there. - Review the result. The request to merge arrives in your Inbox. Every changed record is listed with its values before and after, so the three that do not fit stand out.
- Put back what is wrong, merge the rest. Fix the records on the branch, or revert a commit, then merge. Nothing is lost, and every change stays in the history.
What your team gets
| What you need | What CodeKasten does |
|---|---|
| Mistakes caught before they count | Agents work on a branch a person merges |
| Undo that works | Revert a commit, deleted records included |
| No surprise messages | Calls to other services are held until the merge |
| Decisions kept with people | The agent asks in the Inbox, and the answer is recorded |
| Knowing who did what | Records and commits say "Claude, for Maria" |
| Limits you control | Grants narrow what an agent holds, and you can pause or stop it |
Works alongside your team
People and agents use the same records. Your team works in the app and on the Android app, with chat, an Inbox and reports over the same data. Repeating jobs become workflows, and the agent you already use connects through MCP.
Starting small
A new workspace starts on Free, and nothing is charged yet: see plans. Pick one job that is repetitive, low in risk and easy to check. Let an agent do it on a branch, review it as a team, and see what the review catches. For the thinking behind this, read AI agent approval workflows.