Topic starter
02/08/2026 12:06 pm
TL;DR: Smart systems still fail when local, tacit, and time-sensitive knowledge cannot be centralised, according to Tonic. Bounded autonomy and explicit escalation become essential when agentic workflows must act without full context.
Editorial analysis by NHI Mgmt Group, based on content published by Tonic: “On Building Agents and Agentic Workflows (Part 1)”.
Key questions
Q: Why does dispersed knowledge increase automation risk?
A: Because the most relevant information often lives outside the model, outside the record, and inside human judgement.
Q: How should teams design agentic systems for partial information?
A: Use bounded tasks, explicit stop conditions, and escalation paths for uncertainty.
Q: What breaks when agentic systems assume complete knowledge?
A: They become brittle because centralised state cannot capture the local, tacit, and time-sensitive knowledge that real decisions depend on.
Practitioner guidance
- Define escalation points for uncertain decisions Require AI and workflow systems to route ambiguous or high-impact cases to a human reviewer before they can change access, approve actions, or close incidents.
- Limit agent autonomy to bounded objectives Scope each agent to narrow tasks with explicit inputs, outputs, and stop conditions so it cannot infer broad intent from incomplete context.
- Preserve local operational knowledge in controls Capture frontline exceptions, owner notes, and temporary workarounds in the process that governs decisions, not only in a central record.
Bottom line: Smart systems fail when they are built on the assumption that knowledge can be fully centralised, because the most important context is often local and transient.
Glossary terms: Tacit Knowledge, Context-to-Control Gap, Signal-Based Coordination, Bounded Autonomy
Explore further
View Full Forum → | NHI Foundation Course → | Our Services → | Read the full analysis →
This topic was modified 2 days ago by NHI Mgmt Group
02/08/2026 1:11 pm
Knowledge scarcity is now a governance problem, not just an AI design problem. The article correctly identifies that organisations often overestimate how much can be centralised, normalised, and safely automated. In identity and security programmes, that same mistake appears when teams trust dashboards more than operational context. The practitioner lesson is to govern for decision quality under partial knowledge, not for perfect data completeness.
A few things that frame the scale:
- AI-related credential leaks surged 81.5% year-over-year in 2025, with the surrounding AI infrastructure leaking 5x faster than core LLM providers, according to the State of Secrets Sprawl 2026.
A question worth separating out:
Q: When should human judgment override automated decisions?
A: Whenever the decision depends on context that cannot be reliably encoded, verified, or kept current. If a system is acting on stale state, conflicting ownership, or a situation that changes faster than the control cycle, human review should take precedence before the action is committed.
👉 Read our full editorial: The knowledge problem is why smarter systems still fail