AgentGuard records which configured outcome ran, its policy decision, model, measured token use and cost, and the runtime's local status. A signature protects the recorded metadata from later editing; it does not independently prove a real-world business result or human identity. AgentGuard does not receive prompts or completions; your configured model provider does.
"Applied AI is about jobs to be done, not tokens to be consumed. Token usage may be a reasonable short-term proxy for adoption. But long term, success comes from outcomemaxxing, not tokenmaxxing."
@SierraPlatform · outcomes-based pricing for AI agents, May 27, 2026 · amplified by @saranormous (Conviction)
"A mortgage completed"
The current catalog covers transaction deadlines, lease abstraction, fair-housing copy checks, property operations, screening records, comps assembly, and disclosure packets. It records configured SDK steps; it does not establish that a mortgage closed.
"An order returned"
The implemented return and refund template is read-only: it applies the supplied policy to supplied order facts, recommends a next step, and requires human approval. It never executes a refund. Signing can bind that configured decision metadata.
"A customer saved"
The current catalog can draft source-grounded support replies and chargeback materials. It contains no customer-retention or saved-customer verifier. A customer's application must measure the business result separately.
"A policy comparison drafted"
The insurance catalog includes source-grounded policy variance, ACORD extraction, commission reconciliation, loss-run, renewal, COI, and FNOL templates. Reviewer behavior and receipt signing depend on runtime configuration.
"A basket size increased"
The current catalog has no basket-size or conversion verifier. It can record governed model calls for implemented ecommerce templates, but the merchant must measure any change in order value.
"A contract closed"
The law catalog includes source-grounded clause extraction, chronologies, deposition summaries, billing entries, document assembly, and factual drafting. Its disclaimer keeps a licensed attorney as human of record; it does not close a contract.
"Books closed"
The accounting catalog includes source-grounded categorization, reconciliations, document extraction, trial-balance comparison, return-data preparation, and a month-end close package. Its runtime status does not independently attest that books are closed or compliant.
"Your custom outcome"
Implement an OutcomeTemplate with the outcome-runtime API, or scaffold local BYO outcome YAML with agentguard init. agentguard advisor writes one spend policy and quickstart; it does not execute or certify a business outcome.
The Spend package exports five vertical outcome catalogs and a library API for custom templates. Advisor can be rerun; it writes a local policy.yaml and quickstart from your answers. The current Managed plan exposes the hosted Outcome Builder:
agentguard advisor asks six setup questions and writes one local policy plus quickstart. Its nine explicit profiles affect suggested scope, capability, model pair, caps, and posture; unmatched input uses a generic fallback.
Law firm
Suggests a matter-scoped policy, a compliance posture, a configured primary and fallback model, and spend caps based on the collected answers.
Insurance agency
Suggests a policy-scoped configuration and compliance posture. The generated SpendPolicy does not itself create insurance outcome templates.
Shopify / e-commerce
Suggests an order-scoped policy and standard posture. It does not execute refunds or create task-specific ecommerce routes by itself.
Accounting / bookkeeping firm
Suggests an engagement-scoped policy and compliance posture. Generated records are software logs, not certified or guaranteed compliance evidence.
Real estate / mortgage broker
Suggests a transaction-scoped policy, model pair, and caps. The separate vertical catalog supplies the implemented real-estate templates.
Local services / barbershop
Recognizes listed local-service terms and writes a job-scoped policy. It does not ship a booking adapter or complaint workflow.
Marketing / consulting agency
Writes a campaign-scoped SpendPolicy. Receipt signing remains optional configuration and invoice attachment is the customer's integration.
Software team
Suggests a repository-scoped read-only policy for software terms. Healthcare is the ninth explicit profile; unmatched startup language uses the generic fallback.
Select an implemented vertical catalog or configure an OutcomeTemplate. Templates carry a capability, model route, spend cap, prompt version, inputs, output format, and quality-gate settings.
The outcome runtime asks the in-process Spend guard before each drafter, reviewer, or fallback call it runs. Your configured runner sends prompts directly to your selected provider. AgentGuard does not receive prompts or completions; your configured model provider does.
Signing requires configured Ed25519 keys and decision-log storage. A reviewer with the public key can then verify the signature and hash chain offline or in the browser verifier. That check establishes record integrity, not the truth or legal effect of a business outcome.
Sierra published the outcome-based pricing thesis in December 2024. @saranormous (Conviction GP) amplified it on May 27, 2026. The "tokenmaxxing → outcomesmaxxing" pivot is moving from thesis to default expectation across Applied AI buyers right now.
Outcome-based operations can use outcome-aware infrastructure: per-outcome budgets, capability checks, local quality-gate status, and optional signed technical records. AgentGuard ships those software mechanisms. AgentGuard is patent-pending. The SDK, five vertical catalogs, and AgentGuard Advisor are available under their published package terms.