AI agents with a budget and an approval queue

There are two ways AI shows up in klerc. Molly is the assistant you ask: press ⌘K on any page and she finds data, builds a chart or creates something. AI agents are teammates you configure to work on their own when something happens. Both act through the same actions your team uses, and both rest on the same idea — AI is only useful at work if you can see what it did, control what it costs and stop it changing things behind your back.

Molly: ask, then confirm#

Molly knows your objects and fields, so a question never needs a query language. Ask which renewals are overdue this week, grouped by owner and she runs a real query against your live data, shows the tools she used, and answers with an artifact — a live chart you can switch between renderers, a table or a file — rather than a paragraph.

She can act, too: 165 actions across 19 feature areas, through the same registry the interface and the API use, and gated by the same permissions. But every write waits for you. Ask her to create a follow-up task for each owner, due tomorrow, and she shows a confirmation card first — how many tasks, assigned to whom, due when. Nothing changes until you press Confirm.

Molly runs on the model your organisation chooses — OpenAI, Anthropic, Google or anything on OpenRouter — and you can switch whenever you like.

Agents: teammates with a job description#

An agent is closer to a new hire with a narrow remit. You give it:

• Instructions — what it's for and how to behave. The demo brokerage's renewals assistant, for example, prepares renewal quotes, drafts reminder messages and summarises cascade progress for the renewals team.
• A model, chosen per agent.
• Knowledge — its own slice of the knowledge base. A claims agent sees claim forms and policy wordings; it doesn't need the commission grid.
• Tools — the actions it's allowed to take, such as looking up a policy, drafting a message, creating a task or handing off to a person.
• Playbooks and memories — the procedures it follows, and the facts, preferences and instructions it keeps between conversations.

Before an agent goes live, preview-chat with it to check how it answers.

Woken by events, triaged on purpose#

Agents don't sit waiting for a prompt. They're woken when something they're responsible for happens, look at it, and decide whether to act or leave it. Each agent's overview shows how often it was woken over the last seven days and its triage — how many things it acted on and how many it ignored. Every piece of work becomes a thread, so you can read exactly what the agent saw, decided and did.

A budget you set — and can see#

Every agent runs against a monthly budget. Its overview shows what it has spent this month against that cap, with the tokens behind it, plus cost by day and cost by thread — so an expensive habit shows up as a spike you can trace, not a surprise on an invoice. The demo renewals assistant, for instance, had spent $0.15 of its $50 budget, across 78.1k tokens.

If an agent isn't earning its keep, switch it off from the agents list. Its threads — and everything it did — stay where they are.

An approval queue, not a leap of faith#

Some actions shouldn't happen without a person. When an agent wants to take one, it goes into pending approvals — waiting for a decision — instead of happening on its own. The agents list shows each agent's pending approvals next to its open threads and its cost, so the queue never goes unnoticed.

Together with Molly's confirm step, that's the rule klerc works to: AI proposes; people approve the writes that matter.

Secrets stay secret#

Agents that run code in a sandbox sometimes need an API key or a token. Those live in the Secrets Vault: encrypted, write-only values you can set, rotate or delete. Agents see a secret's name, never its value, and agent outputs are redacted so a secret doesn't leak back out.

A paper trail#

Between threads, approvals and the organisation's audit log — which records logins, record access, changes, exports and admin activity for 45 days, exportable to CSV — there's always an answer to "who changed this, and why?"

A sensible way to start#

1. Pick one narrow job — drafting renewal reminders, say, or summarising claim progress.
2. Give it only the knowledge it needs. Less to read means better answers and lower cost.
3. Choose few tools, and keep writes behind approvals until you trust its judgement.
4. Set a small budget. Raise it when the work justifies it.
5. Preview-chat before switching it on, then read its threads and triage for the first week.

Read more in Molly, AI agents and the Secrets Vault, or see Molly AI & Agents.