Guide

Agent1000 is an AI workforce for national and provincial departments, public entities, municipalities, universities, science councils and SETAs. It takes the administrative work every institution runs on — leave applications, travel claims, budget certificates, contract instructions, the monthly budget report — and does it inside the systems the institution already runs, from Persal to the registry mailbox.

What makes it usable in the public sector is where it stops. An agent prepares; a named official decides. Nothing is captured, sent, published or changed on its own, and every approval is recorded against the person who gave it. That is what lets the decision survive an audit, not just a demo.

Each agent takes one workflow — leave, claims, the budget, contracts — or one standing job such as patching or publishing, and is named for it. Its page in the catalogue says what it connects to, how you reach it, and exactly what a person approves.

Setting up an agent

An agent is set up for your institution, not downloaded. Before it does any work, three things are settled with you:

1. 1The workWhich workflow it takes on, and on what rules: the leave determination, the travel rates and limits in force, the turnaround for each stage.
2. 2The approversThe chain as your institution runs it: line manager and HR for leave; Finance and the supervisor for claims; the delegated official for a contract. The agent sends every binding action to them, and to no one else.
3. 3The accessIts own account in each system it needs — Persal, the document store, the mailbox — with the permissions that job requires and nothing more.

Threads

Each piece of work is a thread: one conversation that holds the request, what the agent did, what it is waiting for and who it is waiting on. Start a new thread for each task rather than running unrelated work through one long conversation.

Keeping the work apart keeps the record readable. A leave application, a travel claim and a budget certificate each have their own thread, so when someone later asks how a claim came to be paid, the trip, the receipts, Finance’s findings and the supervisor’s approval are in one place. A claim sent back for rework stays in the same thread, under the same reference, so the second round can be read against the first.

Memory

An agent learns from its work across all of its threads: the conversations, and the documents and attachments it handled. What it learns in one thread helps in the next — which supervisor signs for which unit, the receipts Finance has queried before, the clauses legal services always asks for — and it can look back at earlier threads when it needs more context.

Memory builds in the background, so there is nothing to manage day to day. It stays with the agent it belongs to; one agent’s memory is not shared with another. And it never stands in for an approval: an agent that has seen a hundred clean claims from the same official still sends the hundred-and-first to Finance.

Sharing an agent

An agent starts with the officials named when it is set up, and more can be added as the work needs them. Everyone working with an agent sees its threads, so a colleague can pick up a task where someone else left it, and a manager can see what is waiting and on whom.

That is how the work is already split. A capturer can prepare a claim for an official who then signs it themselves. A supervisor going on leave can delegate their approvals for a week; the delegate signs “on behalf of”, and the record says so.

Sharing an agent does not share its approvals. Only the officials named as approvers can sign off its work, and every approval is still recorded against the person who gave it.

These are the workflows an institution runs every week, and where an agent fits in each. In every one, the agent does the checking, the routing and the chasing. The signatures stay with the officials who hold them today.

Every task follows the same four steps, in the same order.

1. 1AskSay what you need done, in the words you would use with a colleague. No forms, no scripting, no new system to learn.
2. 2ConnectThe agent works inside the systems the institution already runs: Persal, the mailbox, the document store, the finance system.
3. 3ApproveEvery binding action stops and waits for a named official. The agent says what it wants to do, and to whom it has sent it for a decision.
4. 4RepeatSay “do this every Monday” and the one-off becomes standing work. The approval step comes with it.

What an ask looks like

“Process the leave request from employee 4471, 5 days annual leave, 22 to 26 September.”

“Start my travel claim for the Polokwane trip.”

“Where is the contract instruction for the bursary partnership, and who is holding it?”

You can watch a leave request run, approval gate and all, on the front page.

You reach an agent through channels the institution already uses. Which ones depends on the agent.

An automation is a task that starts on its own: at a set time, or when something arrives or changes. You set one up by saying it in the thread, in plain language.

“Remind me a week before any claim deadline.”

Types of automations

Every agent has one of two approval modes. Neither lets an agent make a binding change on its own.

Who approves what

• VulnWatch: Standing approval to scan and report; an administrator approves each change before anything is patched.
• ContentDesk: The requesting officer approves the preview; nothing goes live until they reply.
• RecordsDesk: A registry clerk confirms the case record before anything is filed against it.
• ReportPack: The responsible manager signs off the pack; the agent assembles it and never submits it.
• SiteWatch: Standing approval to watch and report; an administrator approves any change to a site.
• LeaveDesk: The line manager recommends and an HR practitioner approves; leave is captured on Persal only after both.
• ClaimsDesk: The claimant signs, Finance verifies and the supervisor approves before anything is captured for payment.
• BudgetDesk: The responsibility manager, the budget office and the state accountant each sign a certificate in turn; the responsible manager signs off the monthly report.
• ContractDesk: Legal services reviews every draft, and only the delegated official signs the contract.
• TenderCheck: A supply chain officer will review every flag before it is put to a bidder.
• DeskTriage: A service desk agent will send every reply; the agent will draft and never send.
• PolicyWatch: The compliance officer will decide what action a change requires.
• MinuteDesk: The secretariat will approve the minutes before they are circulated.

Most work in an institution is not a one-off. Once a task has run, one sentence turns it into standing work:

“Do this for every leave request that comes in.”

“Start the budget report on the first working day of every month.”

The approval step is part of the task, so it is part of the standing work too. Every leave request still goes to the line manager, and the monthly report still goes to the responsible manager; neither is ever approved or submitted for them.

An agent signs in as itself: its own account, with its own permissions, in the systems you already run. It never borrows an official’s login.

That keeps the controls you already have in charge. You decide what the account may see and touch, you can see what it touched, and you can take the account away in an afternoon.

A binding action is anything that commits the institution: capturing leave on Persal, verifying or paying a claim, issuing a budget certificate, signing a contract, sending a letter, publishing a page. The agent prepares it; a named official decides. That order is not configurable.

When an agent is unsure — a receipt in someone else’s name, a sick note that does not cover the first day — it stops and says so rather than guessing.

Every approval is logged against the official who gave it, not against the agent. When the Auditor-General, an internal audit committee or a manager asks who authorised a change, there is a person’s name to give them.

Agent1000 is written for the institutions the PFMA and the MFMA already name. The catalogue filters by these classes, so you can see the agents whose work exists in an institution like yours.

Agent1000 is sold and connected through appointed partners. A partner holds one role or both, and the roles answer different questions: who holds your contract, and who touches your systems.

A company that wants to become one can apply as a reseller. Every applicant is vetted before they are appointed.

Before you give a company access to your systems on our behalf, check it on the partner register. Every appointed partner is listed there with its roles, the agents it is certified on, the dates of its appointment, and a certificate number you can quote back to us.

1. Ask the company for its certificate number.
2. Find the company on the register and compare the number.
3. Check that it is certified on the agent it is offering, and that the appointment has not lapsed.

1. 1Describe the workPick the work that eats a morning every week and describe it. You will see which agent comes closest.
2. 2See it runRequest a demo. We show you the agent doing that work, and exactly where it stops for an official.
3. 3Decide who approvesName the officials who will sign off: line managers and HR practitioners, finance officers and supervisors, the responsible manager.
4. 4Connect and startThe agent gets its own account in your systems, connected by your IT or a certified integrator. Its first task waits for its first approval.