Home / Tuvel Control Plane
Every agent on a leash your risk team holds.
Tuvel Control Plane is where authority is decided and recorded: what each agent may do, spend and see, for staff agents and the customer advisor alike. Any of them stops in one click.
Available now What makes the Control Plane different
Autonomy per agent
Each agent holds a level from 1 to 5. Above it, a person decides, and a person approves every step up.
Budgets and limits
Daily and monthly budgets per agent, and caps per payment and per day for customers.
Customer-chosen autonomy
Each customer picks a level for the advisor, inside the ceiling your bank sets per segment.
Approvals
One-button approvals on cards where the work is, with reasons on the record.
Guardrails and policies
What always needs a person, what each agent may see, and baseline banking rules every agent follows.
Model routing per agent
The right model for each agent, through one layer that falls back and prices every call.
Kill switch
Pause any agent in one click. Going over budget stops it too.
Action-level audit
Every action and every decision, with who asked, what ran and why.
- Level
- 1, Observe
- Budget
- $40 a day
- Model
- Set for this agent
- Tools
- 5 granted
- Approver
- Credit officer
Every agent, live.
The live operations view shows your risk team what each agent is doing and has decided, as it happens.
- What every agent is doing now, and what it just decided
- Where the next hand-off is queued
- Pause, demote or override any agent in one click
- Scoped actions, with a blast-radius view before a workflow change goes live
- Maker-checker on changes to connections and rules
- Calls to bank systems in a hash-chained trail, with a signed export to your SIEM
- Loan File OrchestratorRouting file HL-2041 to the credit officerWorking
- Cash Flow and Statement AnalyserIncome matches statements; workings attachedDecided
- Credit Appraisal Note DrafterA deviation on a card for MeeraWaiting
- Sanction to Disbursement TrackerNext hand-off: the pre-disbursement checklistQueued
Earned autonomy.
Every agent holds a level and earns wider autonomy level by level, from its own record. Bank agents start at L1 or L2; a person decides each step up, and a nightly check drops a level when the work slips.
Observe
Reads and reports. Anything it would write, even a draft, waits for a person.
Assist
Reads and drafts on its own. Changes to records other people rely on wait for a person.
Operate
Changes workspace records on its own. Anything that reaches outside the workspace waits.
Represent
Acts outside the workspace on its own, such as emailing a contact. Publishing waits.
Lead
Publishes on its own and hands work to agents below it. Deleting and governance changes always wait.
A results score from its own record
The share of the agent’s counted results that went well, shown once 20 or more are counted and refreshed every night. The score itself never moves a level.
A nightly check on four signals
Over the last 14 days: approval cards dismissed or edited, agent-caused errors, bad ratings from several people and outcomes rated poor. Any one past its line drops the agent one level, at most one a night and never below L1. Provider outages are not counted against it.
Wider autonomy, level by level
After 30 days at its level, with 25 or more decided cards, under 5% dismissed, no level drop and a results score of at least 60, the agent is nominated on a card. A person decides.
A sample sent back to people
About one run in twenty is sent back to a person to grade.
Irreversible actions stay with people
At every level, actions that cannot be undone wait for a person.
Some breaches drop a level at once
Going over budget drops an agent one level straight away. Asking for an action reserved for people, outside an approval card, drops it to L1 and pauses it.
86%Results that went well,
from its 212 counted results
- Level
- 2, Assist
- Days at this level
- 41
- Cards decided
- 184
- Cards dismissed
- 3%
- Agent-caused errors, 14 days
- 1%
- Outcomes rated poor, 14 days
- 0
Levels for customers.
For customers, autonomy takes two keys. The advisor has its own ladder: each customer chooses a level, inside the ceiling your bank sets for their segment, with caps per payment and per day. An action runs on its own only where both allow it.
- Commitments always wait for the customer’s tap
- Loosening a rule needs a one-time code; tightening is instant
- Every rule change is versioned and can be rolled back
Every setting, per agent.
Change one, and exactly one thing changes.
| Setting | What it holds | What it controls |
|---|---|---|
| Level | From L1 to L5 | Whether an action runs now or waits |
| Budget | Daily and monthly caps | How much it may spend |
| Model | Set per agent | Cost and speed, never permissions |
| Skills | Procedures with a priority | How it approaches repeat work |
| Knowledge | Your documents and rules | What it knows before every run |
| Tools | Granted one at a time | What it can reach |
| Approvers | Named people | Who decides above its level |
| Record | Every action | What anyone can review later |
Model routing.
Each agent runs on the model that fits its work, routed through one layer that falls back when a key runs out and prices every call against the agent’s budget.
Audit.
Every turn, tool call, hand-off and decision is recorded, with who asked, what ran and why, for your risk, audit and compliance teams.
Questions.
What does Tuvel Control Plane control?
What each agent may do, spend and see: its level, daily and monthly budgets, the tools and knowledge it is granted, its model, its approvers and the record of every action. It covers staff agents and the customer advisor alike.
How do autonomy levels work for staff agents?
Each agent holds a level from L1, Observe, to L5, Lead. Every action above its level waits for a person. Bank agents start at L1 or L2, and a person decides each step up.
What is earned autonomy?
Autonomy an AI agent earns from its own record. Each Tuvel agent holds a level from 1 to 5. Every night its last 14 days are checked: if people dismiss or edit too many of its approval cards, it causes too many errors, several people rate its work bad or its outcomes are rated poor, it drops one level. After 30 days of clean work at its level, with a results score of 60 or more, it is nominated for the next level on a card, and a person decides.
What is an agent’s results score?
The share of the agent’s counted results that went well, such as approvals, replies and people’s ratings. It shows once at least 20 results are counted, is refreshed every night and on Recompute now, and sits next to the agent’s level, so the person who decides a step up has the evidence in front of them. The score itself never moves a level.
Can we stop an agent at once?
Yes. Any agent can be paused in one click from the control plane, and its next action is refused; resume switches it back on. Going over its budget stops it and drops it one level.
What does the audit record hold?
Every turn, tool call, hand-off and decision, with who asked, what ran and why, for your risk, audit and compliance teams.
Put the controls in front of your risk team.
We walk through every setting, approve a card together and pause an agent live.
Or write to demo@tuvel.co.