Perception
Which authorised sources the agent can observe, how provenance and freshness are checked, and how evidence becomes structured governance state.
How Adnet works
From authorised evidence to a defensible decision and verified improvement.
The method is explicit so management can see where evidence, analysis, authority and accountability change hands.
The operating method
One connected governance record.
The VTGO does not create a duplicate technology estate. It connects the governance objects already held across customer and provider systems and records how evidence supports risks, decisions, actions and outcomes.
- Customer-approved systems remain authoritative.
- Every material claim carries source, owner, date, confidence, access class and review status.
- Contradictions and missing evidence remain visible.
- Decisions create conditions, owners, dates and review obligations.
- Closure requires evidence that the exception is resolved and the intended capability improved.
Exception-to-decision management
Seven controlled stages.
An exception is a material difference between what should be happening and what reliable evidence says is actually happening.
- 01
Observe
Collect authorised evidence and update the current governance state.
- 02
Compare
Test actual conditions against plans, service commitments, architecture, contracts, policies and thresholds.
- 03
Validate
Confirm source reliability, materiality, uncertainty, contradictions and missing information.
- 04
Frame
Prepare realistic options, consequences and a decision-ready recommendation.
- 05
Decide
Authorised people approve, reject, defer, condition or accept risk.
- 06
Follow through
Track conditions, owners, dates, implementation and supporting evidence.
- 07
Verify and learn
Confirm resolution, capability improvement and controlled method learning.
Human-agent architecture
A language model plus governed scaffolding.
Adnet’s design answers five foundational questions. AgentOps then evaluates the complete operating system continuously.
Action
Which tools and reusable governance actions it may use, which records remain provisional, and which commitments are prohibited.
Planning
How work is decomposed, sequenced, re-planned and verified, including when specialist agents or people are called.
Memory
External case memory, client governance memory and reviewed procedural knowledge, all controlled and segregated by customer.
Safety
Least privilege, deterministic access and approval gates, human escalation, immutable records, rollback and kill switches.
AgentOps
Continuous evaluation of evidence completeness, unsupported claims, thresholds, overrides, drift, model and tool changes, and outcomes.
Single before multi-agent
Begin with the least complex architecture that can test the complete workflow.
The pilot uses one bounded coordinating agent. Specialisation is introduced only when it improves expertise, scale, permission separation or independent verification enough to justify additional coordination risk.
A move to multi-agent operation requires evidence that the single agent is a real bottleneck, specialist evaluation is materially better, shared context can be minimised, hand-offs are observable, one authoritative case record remains intact and every conclusion can be attributed and reviewed.
Speed with discipline
Short build cycles. Uncompressed accountability.
Adnet increases learning speed by narrowing scope and authority, not by removing governance.
Internal MVP
Test the minimum end-to-end workflow with no more than five trusted users.
Controlled pilot
Establish baselines, conduct fortnightly evidence reviews and make a scale, revise or stop decision.
Service proof
Operate the cadence with one suitable customer before making a new component a standard commitment.
What does not compress:customer authority and risk acceptance; data access and segregation; material evidence validation; human approval of consequential recommendations; continuous evaluation, incident handling and rollback readiness.