Knowledge & community

Wiki

Patterns, how-tos and reusable components for agentic AI and hyperautomation.

Articles

Pattern: confidence routing auto / HITL / expert: Calibrate confidence on labelled cases; set the auto threshold from the acceptable error rate; add risk rules that always route to experts; review thresholds monthly with override rates.
Pattern: retrieved content is data, not instructions: Wrap retrieved text in delimiters with provenance tags; never let it change tool scopes; filter for instruction-like text; require approval for irreversible actions.
How to add a tool to the MCP registry: Describe the tool, its input/output schema, data classes and side effects; security review; sign and register; bind to agents by allow-list; add contract tests.
How to build an evaluation set: Sample ≥ 200 real cases across variants and edge cases; label expected outcome and evidence; include adversarial cases; freeze a version per release.
How to choose RAG or fine-tuning: See the decision tree on the culture page: RAG for facts and citations, LoRA for style and domain reasoning, DPO for recurring reviewer corrections, rules for limits.
Reusable component: deterministic validator library: IBAN, VAT, dates, amounts, limits, JSON schemas, regex for identifiers; versioned and tested in CI.
Reusable component: HITL console: Proposal, evidence, reason, approve / edit / reject / escalate, four-eyes, seeded control cases, audit trail.
How to run a red-team session: Scope by threat catalogue (TH-01…TH-11); include prompt poisoning in documents; log findings as risks; fix before G2.