All writing
Product · 29 July 2026 · 5 min read
An agent that reads is a report. An agent that writes is a colleague
Read-only agents are safe, useful and stop exactly where the work begins. Here is how we made writing safe instead.
Almost every enterprise AI deployment stops at read. It is the responsible place to stop, and it is also the place where the value stops.
Four properties
A write is safe when it is scoped, reversible, attributable and approved. Anvil will not perform one unless all four hold.
- Scoped — the service account can touch the objects the run declares and nothing else
- Reversible — every write is posted as a batch that can be reversed as a batch
- Attributable — the ledger records the rule, the inputs and the version of both
- Approved — a named human holds the last step
The uncomfortable part
Approval is a queue, and queues rot. So the approval is small: one screen, the diff, the rule, the exception. If approving takes longer than doing it by hand, nobody will, and the agent is worthless.
Next
- 14 August 2026 The 30% your platform never shipped Every ERP covers roughly seventy per cent of what a business needs. The rest is systems engineering, and it is where the money goes.
- 2 July 2026 Two weeks, not two quarters What actually happens in the fourteen days between signing and the first live run.
- 11 June 2026 The ledger is the product We spent longer on the record of what the agent did than on the agent.