← Go back

Guide // MandateGuard

How do you manage spending policies for a team of AI agents?

By the Ingenious AI team · Published 2026-11-14 · Reviewed 2026-11-14

▸ Run the free live demo See pricing

Free covers 1 agent and 1 policy with the full authorize flow, signed receipts and the approval inbox — no card required to start.

Short answer

Manage a team of agents the way you governed one — one agent, one job, one fitted policy — plus three fleet practices: coordinate daily caps so their sum is a consciously accepted worst-case day, keep one approval inbox with a named owner and backup, and run the workspace-level controls that see what individual agents cannot. Duplicate-commitment detection catches cross-agent double-buying, vendor trust aggregates the whole fleet's experience per vendor, and the monthly review reads agents comparatively. Pro and Team provide the multi-agent policies and workspace controls; the method does not change with scale.

What changes when one agent becomes several?

Multi-agent spending policies are a set of MandateGuard policies, one per agent, inside a workspace: each agent has its own key, caps, allowlist and threshold fitted to its job, while workspace-level capabilities — the receipt history, vendor trust computed across the ledger, duplicate-commitment detection and evidence export — operate over all agents together. Pro supports multiple agents and policies; Team raises the limits further.

The unit discipline survives scale

Everything that made one agent governable — identity, fitted policy, receipts, a staffed inbox — remains the unit discipline at five or fifteen agents. What changes is that interactions appear: agents buying overlapping things, daily caps summing past the team's true budget, approval queues mixing five jobs' urgency into one list. Manage the fleet by strengthening the unit and adding exactly three fleet-level practices: coordinated caps, workspace-level duplicate vigilance, and comparative review. Resist the centralising instinct to merge agents into shared identities or one big policy for tidiness; every merger trades away attribution, and attribution is the currency all fleet governance is paid in.

Coordinate caps like a budget, because they are one

Each daily cap is a promise that one agent may spend that much unsupervised in a day. Added together, they are the fleet's worst-case day — a number someone with budget authority should have consciously accepted. Set individual caps from each job's prices as always, then sanity-check the sum against the team's daily risk appetite and adjust the outliers. The same logic applies monthly in aggregate: expected fleet spend is a forecast finance can hold, and the workspace ledger reports actuals in that shape. When a new agent joins, its cap is a budget decision reviewed like any other, not a default copied from a sibling — sibling caps encode sibling jobs.

One inbox, explicit ownership

Approval volume grows with the fleet, and queue design becomes a real job. Keep one inbox and one named owner per team, with a backup; fragmented queues are where approvals go to age. Manage volume at the source: each agent's threshold should be tuned until its escalations are the genuinely judgement-worthy band — a fleet where everything escalates has no supervision, only delay. The webhook can route pending items to the team's channel, and task references must name the job so any approver can judge any agent's card. Periodically sample approvals per agent: an agent whose approvals are always granted in seconds has a threshold problem; an agent whose items wait days has an ownership problem.

Workspace-level controls earn their keep

Duplicate detection and vendor trust are workspace-level by design, and the fleet is where they pay. Duplicates across agents — two jobs buying the same licence, dataset or subscription — are invisible from inside either agent; only the shared ledger sees the pattern, which is why duplication is the signature fleet waste. Vendor trust aggregates every agent's experience with a vendor, so procurement reads the organisation's history rather than one workflow's. Evidence export packages the whole workspace for finance or auditors. These are Pro and Team capabilities, and the honest trigger for upgrading is usually not agent count alone but the first cross-agent incident the free shape cannot see.

Comparative review: read agents against each other

The monthly fleet review reads comparatively. Which agent's denial rate moved, and why? Which agent's approvals have become reflexive? Which allowlist is growing, whose daily cap is never touched (over-provisioned authority) and whose is hit weekly (under-fitted or over-worked)? Comparative reading catches what per-agent reading misses: the fleet's centre of gravity shifting — a research agent quietly becoming a purchasing agent, a provisioning agent's amounts creeping as a provider's prices rise. Policy adjustments then happen per agent, one bound at a time, with the others as the control group. That is a genuinely managerial use of the ledger, and it is only possible because identities were never merged.

Growth path: from one agent to a governed fleet

The practical path runs: one agent on the free plan, fitted and receipted; a second job arrives and justifies Pro, bringing multiple policies, duplicate detection, trust, kill-switch and export; the fleet formalises its rhythms — coordinated caps, one inbox with named backup, monthly comparative review; Team when counts or support needs outgrow Pro. At each step the method is unchanged, which is the point: MandateGuard's fleet story is not a different product at scale, but the same mandate discipline — identity, fitted limits, human judgement at the top band, signed evidence for everything — applied per agent and aggregated per workspace.

Five actions to take now

  1. One agent, one job, one policy — resist shared identities and shared policies
  2. Coordinate daily caps so their sum respects the team budget
  3. Staff the approval inbox with a named owner and a backup per team
  4. Use workspace-level duplicate detection as the cross-agent safety net
  5. Review per-agent receipts monthly; compare agents' shapes, not just totals

Worked example: five agents, one workspace, one cross-agent save

A five-agent operations fleet: provisioning (cap 75, day 300), enrichment (cap 60, day 180), media (cap 50, day 120), research (cap 40, day 110) and purchasing (cap 400, day 900). Coordinated daily caps sum to 1,610 dollars against a team budget comfortable with that ceiling. Each agent's approvals route to the same inbox, owned by the operations lead with the finance partner as backup.

In month two, duplicate detection fires across agents: enrichment and purchasing both moved to buy the same 49-dollar dataset licence a week apart. Neither agent erred — the overlap was organisational, visible only at workspace level. The second purchase is denied from the escalation, the teams agree a data-ownership note, and the fleet's monthly review adds a standing agenda item fed by exactly these cross-agent findings.

Frequently asked questions

Can agents share a policy to save setup time?

They can share a template, not a policy. Shared policies dissolve attribution: receipts no longer answer which agent spent, which is the first question of every review.

How do approvals scale with several agents?

One inbox, one queue discipline. Volume is managed by thresholds per agent — each tuned so its routine band flows — not by adding approvers until nobody owns the queue.

What is the team-level budget control?

The sum of daily caps is the fleet's worst-case day. Set the individual caps with that sum in view; the workspace ledger shows actuals against it.

When do we need Team over Pro?

When agent or policy counts or support needs outgrow Pro. Both carry the multi-agent feature set; Team raises limits and adds priority support.

Can different teams share one workspace?

Possible, but receipts and trust aggregate at workspace level. Separate workspaces per trust boundary keep evidence and approvals cleanly owned.

Keep reading

Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides