Agent/MCP Reliability
Intent → tool selection → arguments → execution → task completion → regression evidence. A connected tool or successful HTTP response is not enough if the business task was not completed correctly.
ATVSP / DIGICREM
Control, verify and measure autonomous AI work.
ATVSP — Autonomous Task–Value–Settlement Protocol — is an emerging DigiCrem architecture for governing what autonomous agents may do, testing whether they complete intended work correctly, controlling the consequences they may create, retaining evidence and measuring machine-scale economics.
Four control layers
From agent intent to accountable business outcome.
ATVSP combines four complementary control problems that become important when AI systems move from answering questions to taking real actions.
Consequence-Bound Authority
Permission to execute an action does not automatically grant authority for every downstream consequence. Sensitive outcomes can require review or be blocked.
Evidence & Verification
Retain the evidence needed to examine what happened, why it happened, which authority applied and whether the intended outcome was achieved.
Machine-Customer Metering
Track usage, quota, runaway loops, cost attribution, machine-readable limits and settlement as agents consume software at machine scale.
Operating chain
One accountable path through autonomous work.
INTENT → AUTHORITY → ACTION → RELIABILITY → CONSEQUENCE → EVIDENCE → VERIFICATION → ACCEPTANCE → ECONOMICS / SETTLEMENT
Consequence-Bound Authority
Action permission ≠ consequence authority.
An agent may be allowed to communicate, negotiate or update a system while still requiring additional authority before creating a financial, contractual, privacy-sensitive or irreversible consequence.
Example: autonomous sales agent
The agent is authorized to negotiate with a customer. The authority boundary changes as the proposed consequence becomes more sensitive.
ALLOW Answer product questions.
ALLOW Offer a pre-approved 3% discount.
REVIEW Propose a 15% discount or non-standard commitment.
BLOCK Expose restricted customer data.
What the control layer records
- Original business intent
- Agent/tool identity and authority scope
- Proposed action and likely consequence
- ALLOW / HUMAN REVIEW / BLOCK decision
- Evidence supporting the decision
- Final outcome and acceptance state
Reliability
A tool call is not the same as task completion.
ATVSP's reliability direction tests the complete path from natural-language intent to business completion and preserves regression evidence as the surrounding model, client, schema or API changes.
TOOL
Selection & arguments
Did the agent choose the correct capability and supply valid, non-hallucinated parameters?
EXECUTION
Result & safety
Did the operation execute successfully within permission and safety boundaries?
OUTCOME
Business completion
Did the intended task actually finish, and can that result be demonstrated with evidence?
Machine-customer economics
Measure autonomous consumption before it becomes invisible cost.
Human-seat pricing and controls may not describe software consumed by autonomous agents. ATVSP's economics direction explores usage, quotas, runaway-loop protection, cost attribution, machine-readable limits, metering and settlement.
AGENT USAGE → QUOTA → RUNAWAY CONTROL → COST ATTRIBUTION → METERING → SETTLEMENT
ATVSP assessment
Map 3–5 real agent workflows.
For action-taking AI systems, DigiCrem can examine representative workflows across intent, existing permissions, actions, possible consequences, authority boundaries and required evidence. The output is a practical ALLOW / REVIEW / BLOCK map and a discussion of reliability and metering gaps.