Cognitive Load: The Two Loads

A judgment failure, not a budget failure

We have priced the conversation carefully — the recurrence, the window, the cost of every carried word. Now watch a failure that no budget explains:

⟨tool_use⟩
search_venues(city = "Molde", capacity = 40)

⟨tool_result⟩
[6 venues, full specifications: floor plans, AV inventories,
catering menus, parking, cancellation policies — 4 pages]

⟨tool_use⟩
get_availability(venue = "all", month = "March")

⟨tool_result⟩
[availability grids for 6 venues × 31 days — 2 pages]

⟨tool_use⟩
get_pricing(venue = "all")

⟨tool_result⟩
[rate cards, seasonal discounts, deposit schedules — 3 pages]

⟨tool_use⟩
book_venue(venue = "Fjordgården Loft", date = "2026-03-12")

⟨tool_result⟩
Booked: Fjordgården Loft, 12 March.

⟨tool_use⟩
search_venues(city = "Molde", capacity = 40,
              accessible = true)

⟨tool_result⟩
2 venues match (of 6; 4 excluded — no step-free access):
[#V2] Kulturhuset Sal B — 60 cap, ground floor
[#V5] Havnesenteret — 45 cap, elevator + ramp

⟨tool_use⟩
book_venue(venue = "V5", date = "2026-03-12")

The reader who always just arrived

Why does a crowded conversation degrade judgment, mechanically? Because of the fact we have leaned on since the beginning: the LLM has no memory. It does not "keep track" of the conversation as it goes — every iteration, the entire conversation is presented fresh, to a reader encountering it for the first time. Look at what one iteration actually is:

Call step 2 the re-orientation tax. It is charged in two currencies at once:

Every word on the wall is paid for TWICE, every iteration:

  once in TOKENS     — the recurrence bills it
  once in JUDGMENT   — a fresh reader must notice it,
                       weigh it, and decide it doesn't
                       matter... without error, every time

The first payment shows up on an invoice.
The second shows up in Example 1a.

Two kinds of load

There is a name for this in the learning sciences. Cognitive load theory — Sweller's framework for how minds handle complex tasks — divides the burden of any task into two parts, and the division transfers to agents almost perfectly:

The classical advice about human learners is: you can't remove intrinsic load — it is the task — so design ruthlessly against extraneous load. For agents, the advice is better than that, and this is the central claim of these lessons: for agents, both loads are designed. Extraneous load is set by what we let onto the wall — that gets its own lesson. But intrinsic load is set by something we chose even earlier.

Intrinsic load is chosen at formulation

The intrinsic difficulty of the agent's job is not the difficulty of the problem. It is the difficulty of the problem as formulated — under the representation we chose and the tool language we granted. Watch the same genuinely hard problem under two formulations:

⟨tool_use⟩
get_enrollments()

⟨tool_result⟩
[14 courses × enrollment lists — 212 students, overlapping]

⟨tool_use⟩
get_enrollments()

⟨tool_result⟩
[summary: 14 courses; 37 conflicting pairs listed by ID]

⟨tool_use⟩
solve_constraints(model = "
  vars: exam[1..14] in {Mon..Fri} × {A, B} × {9:00, 13:00}
  no student-conflict pair on the same day       (37 pairs)
  exam[stats] before exam[econometrics]
  room B unavailable Wednesday
  exams graded by Berg not on Friday
")

⟨tool_result⟩
Solution found (0.2s): [14 assignments, all constraints
satisfied]. Note: solution is not unique; 6 alternatives exist.

⟨tool_use⟩
post_schedule(assignments = "[the solution]")

Hire the model to FORMULATE, not to labor.

Intrinsic load is the difficulty of the CAST, not of the
problem: an agent's best move against a hard problem is to
express it in a form something else can solve exactly.
The judgment stays; the search leaves.

This is a thread we've pulled before, now said at full strength. The framing lesson showed that representation chooses the reasoning — circles are hard in the wrong coordinates. The tool lessons showed the tool set is a hypothesis about how the problem will be solved. Put together: the formulation — representation plus tool language — sets the intrinsic load before the conversation begins. Solvers, databases, compilers, statistics engines: every one is an opportunity to move labor out of the conversation and shrink the agent's job to modeling and judgment.

The two dials

One dial was set before the conversation began; the other is managed while it runs. We've just turned the first. The next lesson builds the instrument for the second — a way to see extraneous load, so it can be engineered instead of felt.