--- title: "The Session Bill · example of The Context Economy · Agentic Atlas" description: "The Session Bill: one agent session's token bill, recomputed six times with one decision changed each time." canonical: "https://agentic-atlas.dev/nodes/session-bill" last-updated: "2026-09-23" --- 1. [Agentic Atlas](https://agentic-atlas.dev/) 2. [Patterns in the Agentic Atlas](https://agentic-atlas.dev/atlas) 3. [The Context Economy](https://agentic-atlas.dev/nodes/context-economy) 4. The Session Bill 1. example of [The Context Economy](https://agentic-atlas.dev/nodes/context-economy) # The Session Bill **How do individual context decisions change a session’s token bill?** In this illustrative session, admitting a skill body later cut its cost about 4.3 times purely from timing. Hook one agent session's token bill, recomputed six times with one decision changed each time Laws & fences - Change one decision at a time, because a combined total cannot show which decision moved the bill by how much. - In this illustrative session, moving research in-house more than doubled the bill. - Tokens kept in context are resent and billed on every later turn, so their cost multiplies. - A project memory note is billed in every window the project spawns until someone edits it, but no bill shows the total. - Caching lowers dollar cost and latency but leaves quality unchanged, since the model still reads every token. When to reach - Reach for this when you accept the context-economy cost rules but cannot tell which ones to design around. - Reach for this when you have read a session's transcript but never seen what it cost. - Reach for this for the ratio and direction of each change, not for absolute token counts. Provenance [session-bill/implementation-the-six-shifts](https://agentic-atlas.dev/nodes/session-bill#implementation-the-six-shifts) · v1.0.11 Addresses atlas_cards session-bill 2. [The Context Economy](https://agentic-atlas.dev/nodes/context-economy) 3. [Statelessness](https://agentic-atlas.dev/nodes/statelessness) 4. [Deferred Context](https://agentic-atlas.dev/nodes/deferred-context) 5. [Baseline Skill](https://agentic-atlas.dev/nodes/baseline-skill) 6. [Subagent Offload](https://agentic-atlas.dev/nodes/subagent-offload) 7. [Context Flow](https://agentic-atlas.dev/nodes/context-flow) 8. [Context Alignment](https://agentic-atlas.dev/nodes/context-alignment) 9. [Cost Relocation](https://agentic-atlas.dev/nodes/cost-relocation) 10. field notes - “the altitude above the single skill (N skills, compounding costs)” - “graded access composed from deferral and distillation” Context AlignmentCost Relocation unfold the map fold the map The card, in place · its connections drawn edges from atlas_links session-bill On this plate [context-the-session](https://agentic-atlas.dev/nodes/session-bill#context-the-session) [problem-signal-the-currency-is-quoted-everywhere-and-metered-nowhere](https://agentic-atlas.dev/nodes/session-bill#problem-signal-the-currency-is-quoted-everywhere-and-metered-nowhere) [choice-a-baseline-bill-then-one-decision-at-a-time](https://agentic-atlas.dev/nodes/session-bill#choice-a-baseline-bill-then-one-decision-at-a-time) [before-the-baseline-bill](https://agentic-atlas.dev/nodes/session-bill#before-the-baseline-bill) [implementation-the-six-shifts](https://agentic-atlas.dev/nodes/session-bill#implementation-the-six-shifts) [result-five-design-deltas-and-one-discount](https://agentic-atlas.dev/nodes/session-bill#result-five-design-deltas-and-one-discount) [verification-what-recomputes-and-what-is-illustrative](https://agentic-atlas.dev/nodes/session-bill#verification-what-recomputes-and-what-is-illustrative) [lessons-what-the-bill-teaches](https://agentic-atlas.dev/nodes/session-bill#lessons-what-the-bill-teaches) [relationships](https://agentic-atlas.dev/nodes/session-bill#relationships) Every section is addressable on its own. Read only the ground you need. ## Context — the session [Permalink to Context — the session section](https://agentic-atlas.dev/nodes/session-bill#context-the-session) A 30-turn [orchestrator](https://agentic-atlas.dev/glossary/orchestrator) session: - **System prompt:** 2,400 tokens - **3 registered skills:** descriptions only, ~60 tokens each → **180 tokens** - **1 MCP server, never used:** 12 [tools](https://agentic-atlas.dev/glossary/tool) × ~300 tokens of schema → **3,600 tokens** (its tools never fire; the session's tool results come from built-ins) - **Conversation:** ~1,000 new tokens/turn (user message + assistant reply + tool results) - **Turn 5:** one skill fires — its **1,700-token body** is [admitted](https://agentic-atlas.dev/glossary/admission) and stays [resident](https://agentic-atlas.dev/glossary/residency) - **Turn 12:** one [subagent](https://agentic-atlas.dev/glossary/subagent) [dispatch](https://agentic-atlas.dev/glossary/dispatch) — 400-token dispatch out, **800-token [return](https://agentic-atlas.dev/glossary/return)** admitted and resident; the subagent burns **45,000 tokens** in its own [window](https://agentic-atlas.dev/glossary/window-context-window) doing the work ## Problem signal — the currency is quoted everywhere and metered nowhere [Permalink to Problem signal — the currency is quoted everywhere and metered nowhere section](https://agentic-atlas.dev/nodes/session-bill#problem-signal-the-currency-is-quoted-everywhere-and-metered-nowhere) The economy's model arrives as laws — the time value of context, the cost classes, whose window pays, cost linearity, the scope value of context, the denomination asymmetry. A reader can accept every one of them and still not know which to design around. Laws don't rank themselves; prices do. And no session prices itself. The transcript is the one artifact a designer reads end to end, and it is exactly the artifact that never shows what it cost to produce — so the sums that decide a design are the sums nobody sees. That is the signal here: **not a session that failed, a session nobody had billed.** ## Choice — a baseline bill, then one decision at a time [Permalink to Choice — a baseline bill, then one decision at a time section](https://agentic-atlas.dev/nodes/session-bill#choice-a-baseline-bill-then-one-decision-at-a-time) This example exists to put a price on [The Context Economy](https://agentic-atlas.dev/nodes/context-economy)'s laws — one shift per law, each shift the shape that law takes once it costs something. The arithmetic is struck in input tokens because that is the denomination that sums; where a decision moves the other two — window occupancy, and the quality that occupancy decays — the shift says so rather than letting the token column stand in for them. The structure is fixed before the numbers are: declare the bill for the session exactly as configured, then hold that session fixed and change **exactly one decision**, so every before→after is attributable to that decision and to nothing else. The rejected alternative is the flattering one — apply all six shifts at once, quote the redesigned session's total, and let the size of the number do the arguing. A cumulative total proves nothing a designer can use: it cannot be decomposed, so no reader can tell which decision earned which part of it, and that attribution is the whole of what a worked example teaches. Six shifts, six laws, one decision each. ## Before — the baseline bill [Permalink to Before — the baseline bill section](https://agentic-atlas.dev/nodes/session-bill#before-the-baseline-bill) [Statelessness](https://agentic-atlas.dev/glossary/statelessness) re-ships everything resident on every pass. Summed over 30 turns (uncached, input side): | Line item | Arithmetic | Input tokens | | --- | --- | --- | | Always-on base (2,400 + 180 + 3,600) | 6,180 × 30 turns | 185k | | Conversation history | ~1,000 × (0+1+…+29) re-shipped | 435k | | Skill body | 1,700 × 26 resident turns | 44k | | Subagent return | 800 × 18 resident turns | 14k | | **Orchestrator total** | | **~680k** | | *(Subagent's own window)* | *boot + 45k, paid once, never by the orchestrator* | *45k* | A ~30k visible transcript, a ~680k bill: the statelessness multiplier is ~23×. Now hold this session fixed and move **one decision at a time**. ## Implementation — the six shifts [Permalink to Implementation — the six shifts section](https://agentic-atlas.dev/nodes/session-bill#implementation-the-six-shifts) Each shift changes one thing about the session above and leaves everything else exactly where it was. Under each, the law it prices and the move that institutionalizes that law. ### Shift 1 — admit the skill body at turn 25 instead of turn 5 Residency drops from 26 turns to 6: the body line falls **44k → 10.2k**. Same artifact, same work done with it — **~4.3× cheaper purely from timing**. *Law: the time value of context — admission cost is integrated over residency, so early real estate is the most expensive.* *The move that institutionalizes it: [Deferred Context](https://agentic-atlas.dev/nodes/deferred-context) — admit just-in-time.* ### Shift 2 — unregister the unused MCP server The 3,600-token schema block stops re-shipping: **−108k, 16% of the entire bill, for zero behavior change** — the session never called those tools. The two skills that never fired cost another 120 × 30 = 3.6k the same way. *Law: the always-on cost class — the standing tax is paid whether or not the artifact ever fires.* *The move: baseline-skill's "earn its slot" ([Baseline Skill](https://agentic-atlas.dev/nodes/baseline-skill)); pruning is *the altitude above the single skill (N skills, compounding costs)*'s job.* ### Shift 3 — do the subagent's research in the orchestrator's window Pull the 45k of reading in-house at turn 12 and it is **not +45k — it is +810k** (45k × 18 remaining turns of residency), more than doubling the bill (~680k → ~1.49M). And dollars are the *survivable* denomination: the orchestrator's resident window jumps ~39k → ~84k, deep into silent decay, for material load-bearing during exactly one task. *Law: whose window pays — misplaced cost doesn't add, it multiplies through residency.* *The move: [Subagent Offload](https://agentic-atlas.dev/nodes/subagent-offload) — the 45k billed once, in a window that was discarded.* ### Shift 4 — run 60 turns instead of 30 Double the turns and the bill does **not** double: always-on doubles (185k → 371k) but history *quadruples* (435k → 1,770k). Total: **~680k → ~2,275k — 2× the session, ~3.3× the bill**, and the growth rate itself is growing. *Law: cost linearity — history is the quadratic-ish row; it re-ships turns 1…N−1 and is the only line that grows by itself.* *The moves: [eviction](https://agentic-atlas.dev/glossary/eviction) by design — compaction, summarization ([Context Flow](https://agentic-atlas.dev/nodes/context-flow)) — and splitting unrelated work into separate sessions to improve [Context Alignment](https://agentic-atlas.dev/nodes/context-alignment).* ### Shift 5 — turn caching on (the odd one out) Each resident token is written once (~1.25×) then read (~0.1×): ~680k full-price equivalents become ~113k — **~6× cheaper, with a workload-dependent latency reduction**. But notice what didn't move: the window still holds every token, attended in full on every pass. **Quality improves 0%.** *Law: the three-denomination asymmetry — caching discounts dollars and latency; nothing buys back quality.* *No move: caching is a platform discount, not context engineering. Shifts 1–4 are the ones only a designer can make — and they are the only ones that touch the third denomination.* ### Shift 6 — add ten "useful" notes to the project memory file Ten lines (~200 tokens) go into `CLAUDE.md`. This session's bill barely moves: 200 × 30 = **+6k, under 1%** — and the subagent's boot quietly carries the same 200 tokens. That invisibility is the trap. The line item recurs in **every window this project ever spawns**: every subagent's boot, every grandchild's, every pass of every future session, until someone edits the file — and no single bill ever shows the sum. It is shift 2 inverted: the unused MCP server at least billed one session visibly enough to get pruned; an inherited line is the same standing tax *sharded across all descendants* — too thin to notice anywhere, too wide to be free. *Law: the scope value of context — required value scales with propagation scope, so the inherited layer carries the system's highest admission bar.* *The move: *graded access composed from deferral and distillation* — the inherited layer keeps the taste, the tree keeps the payload.* ## Result — five design deltas and one discount [Permalink to Result — five design deltas and one discount section](https://agentic-atlas.dev/nodes/session-bill#result-five-design-deltas-and-one-discount) Five design decisions moved this session's bill by −4.3× (timing), −16% (pruning), +810k (window [placement](https://agentic-atlas.dev/glossary/placement)), +1,595k (session length), and +6k-that-is-actually-unbounded (inheritance — the one delta whose true size no session bill can show). The sixth shift is not one of them. Caching is the platform's own discount rather than a design decision: it moved two denominations and could not touch the one that decides whether the work comes out right. ## Verification — what recomputes, and what is illustrative [Permalink to Verification — what recomputes, and what is illustrative section](https://agentic-atlas.dev/nodes/session-bill#verification-what-recomputes-and-what-is-illustrative) Every line of the baseline carries its own arithmetic in the table's middle column, so each row can be recomputed rather than taken on trust, and residency is counted the family's way — tokens × the turns whose inference shipped them, the convention ruled 2026-07-11 in [The Context Economy](https://agentic-atlas.dev/nodes/context-economy) and billed by every worked example in that family. What recomputation does *not* establish is the absolutes. This is a constructed session, not a captured trace: the figures are ratio-shaped by design, and what each shift puts under test is the ratio and its direction, never the token count. So the check that matters is not whether a delta is arithmetically right — it is whether the delta has exactly one cause. A shift is verified when nothing moved with the decision, and that isolation is the property the numbers are arranged to expose. ## Lessons — what the bill teaches [Permalink to Lessons — what the bill teaches section](https://agentic-atlas.dev/nodes/session-bill#lessons-what-the-bill-teaches) Read the deltas, not the totals. Each delta above is predicted by one law of the model and addressed by a named move in the family, and that correspondence — law, price, move — is the payload: the laws are not advisory, they are priced, and the price is what a design decision actually moves. What does not travel is the furniture. The 30 turns, the 12 tool schemas, the 45k of subagent reading are one plausible session's fittings, and another session rearranges every one of them without disturbing a single law. The relationships ledger Evidence-bearing references ## Relationships Every connection keeps the section where it was found. The map above orients; this ledger carries the evidence. ### Outbound references 9 1. in-slice · occurrence 1 [The Context Economy](https://agentic-atlas.dev/nodes/context-economy) what each piece of context costs, in quality, dollars, and latency, and who pays Evidence: [Choice](https://agentic-atlas.dev/nodes/session-bill#choice-a-baseline-bill-then-one-decision-at-a-time) · occurrence 1 2. in-slice · occurrence 1 [Deferred Context](https://agentic-atlas.dev/nodes/deferred-context) keeping a short pointer loaded and fetching bulky context only when a session needs it Evidence: [Implementation](https://agentic-atlas.dev/nodes/session-bill#implementation-the-six-shifts) · occurrence 1 3. in-slice · occurrence 2 [Baseline Skill](https://agentic-atlas.dev/nodes/baseline-skill) the simplest complete skill, and the starting shape later skill patterns build on Evidence: [Implementation](https://agentic-atlas.dev/nodes/session-bill#implementation-the-six-shifts) · occurrence 2 4. undisclosed · occurrence 3 Undisclosed relationship the altitude above the single skill (N skills, compounding costs) Evidence: [Implementation](https://agentic-atlas.dev/nodes/session-bill#implementation-the-six-shifts) · occurrence 3 5. in-slice · occurrence 4 [Subagent Offload](https://agentic-atlas.dev/nodes/subagent-offload) handing work to a subagent so its material never fills the orchestrator's context Evidence: [Implementation](https://agentic-atlas.dev/nodes/session-bill#implementation-the-six-shifts) · occurrence 4 6. in-slice · occurrence 5 [Context Flow](https://agentic-atlas.dev/nodes/context-flow) how context enters, stays in, and leaves a model's window, and how well it's recalled Evidence: [Implementation](https://agentic-atlas.dev/nodes/session-bill#implementation-the-six-shifts) · occurrence 5 7. in-slice · occurrence 6 [Context Alignment](https://agentic-atlas.dev/nodes/context-alignment) how well the available context fits the task at hand Evidence: [Implementation](https://agentic-atlas.dev/nodes/session-bill#implementation-the-six-shifts) · occurrence 6 8. undisclosed · occurrence 7 Undisclosed relationship graded access composed from deferral and distillation Evidence: [Implementation](https://agentic-atlas.dev/nodes/session-bill#implementation-the-six-shifts) · occurrence 7 9. in-slice · occurrence 1 [The Context Economy](https://agentic-atlas.dev/nodes/context-economy) what each piece of context costs, in quality, dollars, and latency, and who pays Evidence: [Verification](https://agentic-atlas.dev/nodes/session-bill#verification-what-recomputes-and-what-is-illustrative) · occurrence 1 ### Inbound references 4 1. in-slice · occurrence 1 [The Context Economy](https://agentic-atlas.dev/nodes/context-economy#examples) one agent session's token bill, recomputed six times with one decision changed each time Evidence: [Examples](https://agentic-atlas.dev/nodes/context-economy#examples) · occurrence 1 2. in-slice · occurrence 1 [Cost Relocation](https://agentic-atlas.dev/nodes/cost-relocation#verification) one agent session's token bill, recomputed six times with one decision changed each time Evidence: [Verification](https://agentic-atlas.dev/nodes/cost-relocation#verification) · occurrence 1 3. in-slice · occurrence 2 [Statelessness](https://agentic-atlas.dev/nodes/statelessness#implications-the-consequence-map) one agent session's token bill, recomputed six times with one decision changed each time Evidence: [Implications](https://agentic-atlas.dev/nodes/statelessness#implications-the-consequence-map) · occurrence 2 4. in-slice · occurrence 1 [Statelessness](https://agentic-atlas.dev/nodes/statelessness#evidence) one agent session's token bill, recomputed six times with one decision changed each time Evidence: [Evidence](https://agentic-atlas.dev/nodes/statelessness#evidence) · occurrence 1 [↑ back to the top](https://agentic-atlas.dev/nodes/session-bill#content) [← the survey](https://agentic-atlas.dev/atlas) Node session-bill · corpus 78c0e17 · Catalog revision e0cb75881244b1a82193ca738b82a0d508dd62e822524ae82873ec79d51dbb61