Agents
Hire an agent. Brief it. Approve its work.
An AI agent in Octahive is a worker with a role, a budget and a person who answers for it. Here is what happens between its brief and your approval.
From hire to approval
Seven things the product does, in the order they happen.
Hire an agent
Give it a name, a role and a monthly credit allowance, and name the person who answers for its work. It joins the roster beside your people.
In the code: An agent is hired like a worker, with a role title, a monthly credit cap and a reviewer who answers for it.
Brief it in the workroom
Describe the job in your own words. The brief, the thread and any files it reads are fenced as untrusted data before the model sees them.
In the code: A person briefs an agent in the workroom; the agent drafts a plan, a run starts only once a person approves that plan, and the person can follow the run and stop the job. Documents, emails, web pages and tool output are fenced as untrusted data before the model reads them.
Read the plan first
The agent drafts a plan: the steps it intends to take, the tools each needs, and what it will not do. Nothing runs yet.
In the code: A person briefs an agent in the workroom; the agent drafts a plan, a run starts only once a person approves that plan, and the person can follow the run and stop the job.
Approve, then watch
A person approves the plan, and one run carries it out. You can follow each step as it happens and stop the job at any point.
In the code: A person briefs an agent in the workroom; the agent drafts a plan, a run starts only once a person approves that plan, and the person can follow the run and stop the job. An agent run’s finished work waits for a person to approve it; agents cannot launch a campaign or send an email.
Every tool call goes through the gateway
The agent holds no credentials. Each call passes through a gateway that checks what this agent may do, and prices and totals come from code.
In the code: Agents act only through a gateway that checks each call; the agent service has no database access or integration tokens (it may hold a model key). An agent drafts a quote; rate cards and code set the price; a person approves.
Spend is on a ledger
Each run draws on the agent’s credit allowance. The credit ledger is append-only, so what was spent, and on what, is always on record.
In the code: Each agent draws on a credit allowance recorded in an append-only credit ledger.
A judge reads the draft before you do
Finished drafts are scored against your agency card and the client’s dossier. Code validates the scores, and a missing answer never counts as a pass.
In the code: A judge scores each draft against the agency card and the client dossier before a person sees it; code validates the scores, and no answer never counts as a pass.
How it works under the hood
What happens between your brief and the books, and where the model is kept out.
Your brief and files
Documents, emails and pages are fenced as untrusted data before an agent reads them.
The agent plans
It works only with the tools it was granted for its role.
The gateway checks every call
Each tool call passes through a gateway that checks what this agent may do.
Code does the maths
Prices, taxes and totals are computed from your rate cards, never by the model.
Changes become proposals
Anything that would reach the outside world waits for a person to approve it.
The ledger records the outcome
Entries are append-only and balance. Agent spend is recorded the same way.
Put the routine on autopilot. Keep the final say.
Get started- A
- T
- P
- L
- S
- N