okikusan-public / articles / AI agents enter the territory code can't write — long-tail × tacit knowledge × tacit thoughts
JA EN
← articles
CONTENTS8 sections
  1. 01Long-tail processes
  2. 02The tacit-knowledge wall
  3. 03A new layer: tacit thoughts
  4. 04Layers of knowledge
  5. 05Why only agents reach this
  6. 06Real-world agent map
  7. 07Obsidian × AI practice
  8. 08Long-tail DX
  9. Summary
TACIT × LONGTAIL / 2026

AI agents enter the territory code can't write — long-tail × tacit knowledge × tacit thoughts

Automation has always stopped at explicit knowledge. RPA, SaaS, core systems — they all only swallow work that can be flowcharted. AI agents are the first to descend into the tacit, and into the layer just before that — what I'm calling tacit thoughts: the situation-aware territory.

Connecting Nikkei xTech's April 2026 "long-tail × AI" piece with Polanyi's tacit knowledge, Nonaka's SECI, Karpathy's LLM Wiki, and "tacit thoughts" (a coinage) into one layered model.

Long-tail processes Tacit knowledge Tacit thoughts AI agent Polanyi SECI DX 2026.05.19 · 8 min read
FIG.0 — LAYERS OF KNOWLEDGE
▸ waterline (verbalisation) // Layers of knowledge — the order AI agents descend through ▸ EXPLICIT Explicit knowledge Manuals / SOPs / already in code What conventional automation could reach ▸ TACIT Tacit knowledge Experts' judgement, knack, exception handling Where long-tail work was sunk ▸ SUBTACIT Tacit thoughts Wobbles / hypotheses / hunches / direction of curiosity Raw material that exists only inside the person descending // "What code can write" is only the tip. AI agents descend into the lower layers.
Above the waterline = explicit. Below it = tacit. Sediment at the bottom = tacit thoughts. "What code can write" is only the tip. AI agents descend into the lower layers.
▍ THE PROMISE

The frontier of automation has moved from explicit knowledge to situation-aware work. Only LLM agents can descend into the tacit and sub-tacit — which is why long-tail processes finally crack open and why individual tacit thoughts become usable for the first time.

▍ SOURCES
▍ FROM THE FIELD — from a 5/17 daily note

Compiled from two Vault notes: 5/17 "Long-tail processes × AI agents" and 5/18 "Obsidian × Hermes — tacit thoughts as your own twin." A team of one or two at a corporation and a single person's head share the same structural problem (the tacit layer dwarfs the rest), and both become tractable only with AI agents.

▍ TL;DR
§ 01 LONGTAIL

Long-tail processes — the seam of 1–few owners

Large enterprises follow a power-law distribution of work. The head — hundreds to thousands of people doing the same thing (expenses, ordering, HR requests) — has been covered by core systems and RPA for years. The tail, on the other hand, hosts vast variety performed by only one or two people each. In April 2026, Nikkei xTech named this "long-tail processes."

FIG.1 — LONGTAIL DISTRIBUTION
// Headcount-by-process distribution inside a large enterprise headcount → process variety → High-volume routine Core systems / RPA Mid-volume SaaS / business apps ▍ Long-tail processes 1-few owners · highly varied · person-dependent Was: hard to systemise / Now: AI agents can reach here // The tail is the "tacit-knowledge sink" — deterministic code never reached it
Headcount distribution of corporate processes. Head = core / RPA, mid = SaaS, tail = 1–few owners. The tail is where tacit knowledge piles up — deterministic code never reached it.

Examples Nikkei lists:

The common pattern is clear: "not a process with hundreds of people," "flow differs every time by team/case," "the rules live only in the operator's head." So RPA / SaaS / core systems can't justify the ROI.

▍ In one line

Long-tail processes = 1–few owners × highly varied × person-dependent. They were dropped from systemisation because they were "important but impossible to optimise as one thing" — the last automation frontier inside the enterprise.

§ 02 TACIT

The tacit-knowledge wall

The reason long-tail processes resisted automation isn't low volume. It's that their substance is almost entirely tacit knowledge. That's the real wall.

>2-1Polanyi's tacit knowledge

The philosopher Michael Polanyi (1958) gave us "We know more than we can tell." People know beyond what they can verbalise. Riding a bike, recognising faces, an expert craftsperson sensing "the curvature is off" — none of these are fully explainable by the person doing them. Polanyi called this tacit knowledge.

>2-2Nonaka's SECI and "externalisation"

Ikujiro Nonaka's SECI model (1995) frames organisational knowledge creation as a loop: Socialisation → Externalisation → Combination → Internalisation. The hardest step is Externalisation — turning individual tacit knowledge into words/diagrams others can use. This is exactly where long-tail automation gets stuck.

>2-3Long-tail = a tacit-knowledge sink

The fewer the operators, the more person-dependent the work becomes. The more person-dependent, the more judgement criteria, exception handling, "last time we did it this way so this time too" local rules pile up as tacit knowledge inside one head. Nikkei calls this out as the biggest wall:

Interviews with operators are needed to verbalise tacit knowledge. If extraction is shallow, the AI behaves incorrectly and can't handle exceptions.

In other words: the long-tail layer is sunk beneath a tacit-knowledge ocean. To put it in code at all, you have to surface the tacit first.

§ 03 SUBTACIT

A new layer in front of that: "tacit thoughts"

There's one layer below this. The layer just before tacit knowledge. I'll call it tacit thoughts (a coinage).

>3-1Definition

Tacit knowledge is "judgement that has solidified inside an expert, even if they can't explain it." Tacit thoughts are "raw material before it solidifies." In SECI terms, this is the still-circulating cognitive layer inside one person, sitting before tacit knowledge.

>3-2Examples

▍ Why call them "tacit thoughts"?

If tacit knowledge is "knowledge before it can be told," tacit thoughts are "thinking before it becomes knowledge." They sit upstream of Polanyi, are more fragile, and exist in vastly greater volume. Everyone generates them daily — yet until now, there was nowhere to record them and no one to use them.

§ 04 LAYERS

The layered model — the order AI descends through

Stack the three, and you get a layered model of knowledge:

Automation history has descended through this stack:

§ 05 WHY-AGENT

Why only AI agents can descend into the tacit

The crux is simple: code is deterministic, agents are situation-responsive.

FIG.2 — CODE × AGENT
// The frontier of automation has shifted to "situation-aware" work ▸ CODE / RPA / SaaS Deterministic automation if A → do X / if B → do Y [×] Explicit knowledge only [×] Ceiling = what you wrote ahead of time [×] Breaks on edge cases [×] Can't reach the long tail ▸ LLM AGENT Situation-responsive autonomy read → judge → act (every time) [✓] Descends into tacit / sub-tacit [✓] Improvises on edge cases [✓] Helps verbalise tacit [✓] Lands in the long tail SHIFT // On the same task, agents act differently every time depending on the situation
Left: deterministic automation (ceiling = what you wrote ahead of time). Right: situation-responsive autonomy (read → judge → act, every time). Even on the same task, agents act differently each time depending on context.

>5-1Code = deterministic

RPA, SaaS, classical scripts — all the same principle. You write "if A then X" and "if B then Y" ahead of time. Inside the branch net you wrote, it runs perfectly. Step outside the net, and it breaks. Long-tail processes are "outside the net" by nature, so deterministic approaches are structurally wrong for them.

>5-2Agents = situation-responsive

LLM-based agents work differently. Every time, they read the situation, judge, and act. Same task, different input → a different route. If something exceptional comes in, the agent doesn't force a fit — it asks for more info. Even with no formalised judgement criteria, it can infer from natural-language context and still act. This is the property we call "situation awareness."

>5-3Tacit and sub-tacit are "language work"

Tacit knowledge and tacit thoughts can't be captured by code, but they can be expressed as nuance in natural language. LLMs have the rare property of handling linguistic ambiguity directly. From incomplete descriptions like "in cases like this I usually do X" or "something feels off" or "probably this one," an agent can compose a situation-aware response.

▍ In one line

The frontier of automation has moved from "making explicit knowledge efficient" to "taking over situational improvisation." The territory code can't write is exactly where AI agents are at their best.

§ 06 EXAMPLES

A map of agents already shipping — present tense, not future tense

To keep this from staying abstract, here are working systems as of May 2026. Map them on two axes — individual / business × tacit / sub-tacit — and all four quadrants already have real agents.

FIG.4 — REAL AGENTS DESCENDING INTO TACIT
// Real agents mapped on individual/business × tacit/sub-tacit Individual Business Tacit knowledge ↓ Tacit thoughts ↓ ▍ Individual × tacit Claude Code / Codex Reproduces an expert's coding judgement Cursor Substitutes code-context intuition ▍ Individual × sub-tacit Hermes Agent (OSS) Twin that reads tacit thoughts in Obsidian ChatGPT / Claude.ai Verbalises wobbles via dialogue ▍ Business × tacit Daikin's tacit-knowledge AI Manufacturing / NEDO top prize + Agent prize Misumi's tech-support AI ~90% accuracy (veteran reproduction) ▍ Business × sub-tacit Jitera Turns tribal knowledge into living context "between the lines" Things rarely written down, made explicit // Real systems exist in every quadrant. The tacit territory is a present-tense market, not a future one.
Four-quadrant map: individual × business × tacit × sub-tacit. Every quadrant already has shipping systems. The tacit territory is a present-tense market, not a future one.

>6-1Business × tacit — Daikin / Misumi

Both exemplify "business-specific × externalisation of tacit knowledge." The "interview operators to surface tacit knowledge" step Nikkei pointed to is being performed continuously by AI embedded inside the business.

>6-2Individual × tacit — Claude Code / Cursor

>6-3Individual × sub-tacit — Hermes Agent / ChatGPT・Claude

>6-4Business × sub-tacit — Jitera

Compared with Daikin / Misumi, which AI-ify Polanyi-style tacit knowledge, Jitera steps directly into the phase that is "not yet written, not yet formed" — making it the representative example of the business × sub-tacit quadrant.

▍ All four quadrants are filled

In 2023, every quadrant was "future talk." As of May 2026, every quadrant has a commercial or OSS example shipped. The tacit territory has moved from debate-and-hypothesis to the front line of implementation competition.

§ 07 PRACTICE

Personal practice — Obsidian × AI as your twin

At the individual level, the pattern is pile tacit thoughts roughly into Obsidian × an AI agent reads them as your twin. The intellectual foundation of the setup in the Hermes Agent piece sits here.

FIG.3 — TACIT-THOUGHTS PIPELINE
// Tacit-thoughts pipeline — pile / read / act ▸ WRITE▸ POOL▸ READ▸ ACT Scribble rough Human / daily No structure / no tags Sub-tacit pool Obsidian Vault Wobbles / hypotheses / hunches Read & recompose AI agent Different every time, by context Output / action Posts · decisions · requests Situationally responsive // Tacit thoughts become "actionable" only by going through the agent
Scribble → Obsidian piles it → AI reads it and recomposes by situation → output / action. Tacit thoughts become "actionable" only by going through the agent.

>7-1The value of "not organising"

Tacit thoughts die the moment you organise them. Tagging, formatting, headings — all of these strip the raw material before it solidifies. So the right move for daily/ is leave it rough. Wobbles, contradictions, half-formed hypotheses — keep them all. Demand polish, and the will to input dies entirely.

>7-2Two-layer setup: daily / knowledge

The structure of Karpathy's LLM Wiki, applied directly. The division is clear: human = the device that generates tacit thoughts; AI = the device that organises and uses them.

>7-3The role of a resident agent (Hermes)

A resident agent (e.g. Hermes) reads the Vault daily and lifts "what's usable now" from tacit thoughts and recomposes it. Ping it from your phone with "summarise that thing" — Hermes reads your Vault's tacit thoughts and answers. That's the moment tacit thoughts first become "things you can call up and use."

▍ "Obsidian = where thinking accumulates / AI = the entity that uses it"

Once this division works, AI stops being a single-shot chat partner and becomes a continuous partner with your own context. An AI that can read your tacit thoughts knows your judgement, habits, and curiosity — enough to be worth calling your twin.

§ 08 BUSINESS

Long-tail DX — from interview support to tacit-thoughts collection

The same model maps directly onto enterprise DX. The Nikkei piece's "AI-assisted tacit-knowledge interviews" get even sharper when you descend further to the tacit-thoughts layer.

>8-1Old DX: harvest the head only

RPA / SaaS / core-system rollouts only harvested head-side hundreds-of-people processes and stopped there. The tail was left alone because the ROI didn't pencil. On the ground, this kept producing the same complaint: "we rolled it out and it still isn't easier."

>8-2Surfacing tacit knowledge (short term)

Stage one is LLM-assisted tacit-knowledge interviews. The operator and an AI talk back-and-forth, surfacing judgement criteria, exception handling, and knack. AI takes over the externalisation step that Nikkei called "the biggest wall." In SECI terms: accelerating Externalisation.

>8-3Collecting tacit thoughts (mid term)

Go one step further, and the operator's wobbles, hunches, and hypotheses — i.e. tacit thoughts — become collectable too. Until now, there was nowhere to record this as "business know-how." In the agent era, it makes sense to have an "operator tacit-thoughts pool" in an Obsidian-like place, with an AI using it. Only then do quality of exception handling and quality of handover finally lift.

>8-4Delegated execution by resident agents (long term)

Once tacit and tacit-thought layers accumulate on the AI side, processes the operator no longer has to run personally start expanding. Unlike RPA, agents judge by situation each time, so they can absorb the "slightly different every time" nature of the long tail. This step can overtake a decade of deterministic-code accumulation at once.

▍ The Applied Engineer / FDE role

The value of Applied Engineer / FDE roles (covered in a separate post) ties in directly here. Going down from the field's tacit knowledge into tacit thoughts and pulling verbalisation out of people is something neither an upstream consultant nor a pure engineer can do alone. "The person who runs all three layers end-to-end" is the lead role of long-tail × AI-era DX.

▍ THE WORLDVIEW — situational improvisation is the agent's home

The territory code can't write is the agent's main stage

Automation has descended layer by layer: explicit → tacit → tacit thoughts. Deterministic code (RPA / SaaS) can't climb past the head's explicit-knowledge ceiling. Only LLM agents act situationally, which is why they descend into tacit and sub-tacit.

At the individual level, the pattern is pile tacit thoughts in Obsidian; an AI twin reads and uses them. At the enterprise level, long-tail processes finally become automatable. The same structural problem is being solved on both fronts at once.

  • Humans = devices that generate tacit thoughts (rough, ungroomed, in volume)
  • AI = the device that reads and uses them (recomposed by context every time)
  • That division cracks open the long tail and grows a personal twin

From "automate it in code" to "hand the uncodifiable to AI." Not a continuation of efficiency work — a qualitative change in what automation can touch. Long-tail processes opening up and personal twins becoming real share the same root.