“I shall reconsider human knowledge by starting from the fact that we can know more than we can tell.”Michael Polanyi · The Tacit Dimension · 1966 · The Tacit Dimension (1966), cap. 1 «Tacit Knowing»
What can be told gets coded; what is known but cannot be told goes to the agent, with approval.
The 2026 hybrid pattern: an orchestration layer (the agent) plans and handles exceptions; deterministic executors (RPA bots, APIs, scripts) do the steps; RPA vendors reposition their bots as the "execution layer". Design: every step is classified explicit (a rule exists, so code) or tacit (people say "it depends", so an agent with approval); the agent never touches the system of record directly, it calls an executor with validation; exceptions go to a human queue and become training data for the next rule. Polanyi's point: tacit knowledge does not become explicit because you ask for it; the agent takes it on through examples (resolved cases in the prompt, or fine-tuning, lesson 14). The measure: the exception rate should fall over time as rules are extracted.
Polanyi teaches that we know more than we can say. The next step: how do you recognize a tacit step in practice? You don't look at the procedure, you look at what people ask. If a colleague, before acting, says "it depends on the client" or "it depends on the season", there is no rule to write there. Example: an accountant approves an unusual invoice. The procedure only says "check the supplier". He knows, without explaining, that for this supplier he also checks the payment history. That is the signal. So the first step of hybrid design is not technical but observational: sit next to someone and note the moments when they hesitate. Hesitation is the fingerprint of tacit knowledge. But what do you do with these moments once collected?
Passage 2: once you have collected the hesitation moments, you sort them into two piles. The simple rule: if a step can be fully described, it becomes code or an RPA robot; if the answer is "it depends," it stays with the agent, with human approval. Example: for invoices, matching the amount to the purchase order is fully describable, so a script handles it. Approving an invoice from an old supplier with a changed price is tacit, so the agent handles it. The detail that matters: the agent never touches the system of record directly. It calls an executor, and the executor validates. So if the agent errs, the error is caught at validation, not in accounting. It is like an apprentice who does not sign alone, but prepares the file the master signs off. The open question: what happens to exceptions neither the agent nor the human has seen before?
Why it matters The part nobody could write into the procedure is the part worth an agent.