Guide // MandateGuard
Why use a task-specific mandate? One task, one envelope
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
A task-specific mandate grants spending authority for one job rather than for an agent's whole life: this task, this maximum total, these vendors, this period, these approval conditions. MandateGuard implements it as the spend envelope, evaluated by the real policy engine with decisions in the shared receipt chain. The alternative — stretching the standing policy for every exceptional job — permanently inflates caps and allowlists with permissions nobody remembers granting. Envelopes keep the standing policy boring, make exceptional authority expire with the task, and give clients and auditors a self-contained artefact per job.
What is a task-specific mandate?
A task-specific mandate is authority scoped to one task rather than to an agent's whole operation. MandateGuard implements it as a spend envelope: the task's name and intent, a maximum total, approved vendors, a period and approval conditions, evaluated by the policy engine with decisions in the shared receipt chain. The agent's standing policy continues to govern its ordinary work; the envelope governs the exceptional job laid over it.
The scoping principle
Every authorisation system eventually learns the same lesson: authority should be as narrow as the work it enables. Standing policies are necessarily broad — they cover an agent's whole life, all its ordinary days. Task-specific mandates apply least-privilege thinking to money: this job may spend this much, with these vendors, during these dates, under these conditions — and nothing in that grant leaks into the agent's other work or survives the job's end. The practical dividend is not just safety. Narrow grants are easier to approve, because the approver judges a bounded story instead of an open-ended one, and easier to audit, because the mandate and the task share a name.
What breaks when you stretch a standing policy instead
The alternative — loosening the standing policy for each exceptional job — fails predictably. Caps rise for a project and stay risen. A project vendor joins the allowlist and is still there, payable, a year later. The approval threshold moves to accommodate one big legitimate purchase and thereafter supervises nothing. Each change is rational in the moment and corrosive in aggregate, and the policy quietly stops describing the agent's real authority. Post-incident reviews then discover permissions nobody remembers granting. Task-specific mandates interrupt that ratchet: exceptional authority lives in an instrument with an expiry, and the standing policy's integrity survives contact with exceptional work.
The envelope as the implementation
MandateGuard's spend envelope carries the task mandate's fields: intended task, agent, maximum total cost, approved vendors, optional period start and end, approval conditions and an optional per-spend approval level. Defining one is a short form, not a project. Evaluation is by the real engine: a spend request against the envelope is computed against its terms and presented on the reviewer card with the verdict — within mandate, needs approval or outside mandate. Decisions write into the workspace's hash-chained receipt trail, so the task's story inherits the same tamper-evidence as all other spend. The envelope gets an ID (env_…) that requests reference, keeping task attribution explicit.
Choosing the task boundary
The craft lies in where the task boundary is drawn. Too broad (Q3 operations) and the envelope becomes a shadow policy with worse review habits. Too narrow (buy one dataset) and operators drown in envelope administration for work the standing policy already fits. The sweet spot is the unit of work that has its own budget conversation: a client project, a sprint, a migration, a campaign. If a piece of work would get its own line in a budget review or its own invoice grouping, it is a candidate for its own mandate. If it is Tuesday's ordinary operations, the standing policy is the right home and an envelope is ceremony.
Expiry as a feature
An envelope's period is not metadata; it is the mechanism that returns authority automatically. When the period ends, requests against the envelope stop fitting it — no cleanup ticket, no forgotten allowlist entry, no permission archaeology at the next audit. Teams should still hold a closing review: compare the envelope's maximum with its actual spend, read its decisions, and carry the lessons (a vendor to promote to the standing allowlist deliberately, a condition that proved unnecessary) into the next task's mandate. Expiry plus review is the full lifecycle: authority granted narrowly, exercised transparently, ended automatically and examined honestly.
Task mandates across a fleet
As fleets grow, task mandates become the unit of delegation. A lead can grant a team's agent an envelope for a project without touching the fleet's standing policies; finance can require envelopes for any client-billable agent work; auditors can sample closed envelopes as self-contained case files. The pattern also clarifies accountability: the person who defined the envelope is the person who authorised its terms, and their name belongs in the envelope's creation record and the retrospective. Standing policies answer who watches the agents; task mandates answer who authorised this work — and mature operations need both answers on demand.
Five actions to take now
- Fork authority per task when the task's budget, vendors or approver differ from the agent's norm
- State the task as a deliverable with an end, not an activity without one
- Fund the envelope from the task's own estimate or client budget
- Let the envelope expire with its period — renewals are new decisions
- Review closed envelopes against outcomes at the task retrospective
Worked example: a client project that never touched the standing policy
An agency runs one operations agent under a modest standing policy. A client onboarding project arrives with a 1,200-dollar tooling budget: two vendors the standing policy does not allow, a three-week window and a client requirement that any single spend over 300 dollars be approved by the agency's lead. Instead of rewriting the policy, the agency defines an envelope for the onboarding task with exactly those terms.
The project spends 1,040 dollars inside the envelope; one 340-dollar request pauses for the lead's approval on the reviewer card. When the project closes, the envelope's period ends and the two project vendors lose their authority automatically — they were never in the standing allowlist. The client's invoice review reads the envelope and its receipts directly.
Frequently asked questions
Can an agent work under several envelopes at once?
Yes — different tasks can carry different envelopes concurrently. Requests are evaluated against the envelope they claim, and receipts keep the stories separate.
Does an envelope override the agent's policy caps?
An envelope is its own mandate surface with its own maximum and conditions, evaluated by the engine. Design the pair consciously: standing work under policy, task work under its envelope.
What if the task outgrows its envelope?
That is a budget conversation, conducted deliberately: close and redefine, or define a successor envelope. Quietly stretching defeats the instrument.
Who approves envelope spends?
The reviewer working the envelope's card, applying its written conditions — the same human-in-the-loop discipline as the inbox, scoped to the task.
Are envelopes suitable for client billing?
They are the natural fit: the mandate, decisions and totals for one client's task form one auditable artefact.
Keep reading
- What is a spend envelope for an AI agent?
- Spend envelope vs standing policy: which should you use?
- How do you define a spend envelope, step by step?
Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides