Agent patterns

Autonomy level
Maturity
Evidence status
Primary solution area

16 pages

For low-risk, internal-only actions, requiring human review on every single instance is itself a bottleneck research found production deployments now avoid for…
KCS's own core insight: knowledge captured as a byproduct of solving a problem is more complete and more likely to happen than knowledge authored separately…
Manual spot-checking of every completion does not scale to fulfilment-centre volumes; automated verification with a human override path does.
Manually configuring each new connector is the single largest driver of onboarding time for a new tenant or a newly acquired subsidiary.
Drift between scheduled reviews (CAP-113) is invisible until the next cycle catches it; a continuously running rule set closes that gap without waiting for a…
The action could not have been requested because nobody knew there was anything to request until the contradiction itself was found — the domain's own origin…
Two live, both-currently-published sources can disagree with neither one being obviously wrong — detecting this requires comparing sources against each other,…
The value is entirely in catching the duplicate before re-diagnosis starts, not after — a match surfaced after the fact saves nothing.
A static question list cannot follow up on what an expert actually says — eliciting genuinely tacit knowledge requires a conversation that adapts in real time.
A fluent, confidently-worded answer with a broken citation is worse than a visible failure — this check exists to catch that specific failure mode before it…
Bridges the gap between eliciting tacit knowledge conversationally and the structured format a safety-critical procedure requires.
A single ranking strategy tuned for one query type performs worse on the others — routing first improves the quality of everything downstream.
Manually identifying every piece of identifying detail across a full engagement's content is slow and error-prone; a first-pass proposal focuses the human…
Capture without structure just relocates the unstructured blob, the specific failure EX-118 works through — structuring has to happen immediately, not as a…
Succession risk is easy to see in hindsight and easy to miss proactively — systematic scanning surfaces it before a departure forces the issue.
Governance evaluated after the fact cannot prevent the action it is meant to govern — it has to sit in the path of the call itself.