Instrucciones de origen · Vista previa de solo lectura
name
context-economics-for-agent-swarms
description
Treat tokens as a coding swarm's cost-of-goods-sold AND its legibility engine. Covers per-agent context budgeting, compaction strategies (in-context summarization, structured note-taking, sub-agent isolation), the shared digest as compaction-for-humans- and-agents, the context-degradation cascade (lost-in-the-middle, context rot, recursive- summarization collapse), and mechanism design for metering token spend across a fleet. Activate on: "token budget", "context budget per agent", "compaction strategy", "digest / briefing as compression", "context window degradation / context rot", "summarization collapse", "COGS for agents", "metering agent token spend", "pheromone decay as forgetting", "/context-economics-for-agent-swarms". NOT for: memory storage architecture (use always-on-agent-architecture), single-prompt wording (use prompt-engineer), pure mechanism-design proofs (use nisan-et-al-2007-algorithmic-game-theory).
license
Apache-2.0
metadata
{"category":"AI & Agents","tags":["context-economics","token-budget","compaction","legibility","digest","agent-swarms","cogs"],"pairs-with":[{"skill":"always-on-agent-inputs","reason":"Per-tier context budgeting + retrieval strategy this skill prices"},{"skill":"wang-et-al-2025-tdag","reason":"Context precision vs bloat and the cascading-failure model"},{"skill":"nisan-et-al-2007-algorithmic-game-theory","reason":"Mechanism design for metering and charging token spend"},{"skill":"episodic-memory-algorithms","reason":"Promotion of transient turns into durable compressed episodes"}]}
io-contract
{"kind":"deliverable","produces":[{"kind":"design-doc","description":"Context budget allocation, caching, and compaction strategy across a swarm with cost accounting"}]}
Context Economics for Agent Swarms
The two-sentence thesis. In a fleet of coding agents, tokens are simultaneously
the bill and the map. They are the cost-of-goods-sold (COGS) — the metered,
per-task variable cost any economy must account for — and they are the legibility
mechanism: the only way a human (or another agent) ever sees what a swarm did is a
compaction of its raw trajectory. The digest IS compaction. So the same engineering
choice — what to keep, what to drop — is at once a cost-control decision and a
truth-telling decision. Optimize one without the other and you ship a fleet that is
either bankrupt or illegible.
This skill is for deciding how many tokens each agent gets, how that context is
compacted as it fills, how the swarm's shared digest is produced, and how to detect
the failure cascade when compaction lies.
CORE FRAME — three identities of one resource
Identity
What a token is here
Who pays attention
The optimization target
COGS
a metered variable cost charged per task
the operator / the market
minimize $ per landed unit of work
Working memory
a slot in a finite attention budget
the agent mid-task
maximize signal density before degradation
Legibility lens
a unit of the digest someone else reads
the human + successor agents
maximize recall-of-truth, zoom preserved
Legibility-with-zoom (the cardinal rule). Every digest is a lens onto the real
artifact, never a replacement for it. Over-flattening — a summary you cannot zoom
back from to the source — is the failure mode, not the goal. (This is Scott's warning
about high-modernist over-legibility crushing local knowledge, applied to compaction.)
DECISION POINTS
1. Per-agent context budget allocation
What is this agent's role on the task?
├─ Orchestrator / planner
│ └─ SMALL working budget, LARGE digest-read budget.
│ It should hold the plan + sub-agent summaries, NOT raw tool output.
│ Budget: ~system 4K + plan 4K + N×(1–2K sub-agent digests).
├─ Deep worker (edits files, runs tools)
│ └─ LARGE working budget, but CLEAN window (sub-agent isolation).
│ Give it ≤3–5 always-loaded tools; discover the rest. Return a
│ 1–2K distilled summary upward, NOT its transcript.
└─ Reviewer / auditor
└─ MEDIUM budget anchored on the DIFF + the obligation, not history.
Re-derive from artifacts (git, claims) over scrollback.
Rule of thumb: budget to the effective context, not the advertised window. Empirically
the window where models stay within ~85% of short-context accuracy is far smaller than
the marketing number (often single-digit-thousands of tokens), so packing to 200K is
buying degradation, not capability.
2. Which compaction strategy when context fills
Two families of move, both non-destructive when done right.Macro-compaction
operates on message ranges: lay a non-destructive overlay over a span of cold
history — a pointer overlay when the range is reconstructable, a summary overlay when it
isn't — and never delete the original, so you can always zoom back. When you re-summarize,
pass the old summary back in view and update it rather than re-deriving from scratch
(that iterative discipline is what stops recursive-summarization collapse — see FAILURE
MODES). Micro-compaction is the cheap read-time pass underneath: truncate individual
aged tool outputs (keep the last ~4 intact, replace oversized rows with previews)
continuously, long before you reach for a macro pass.
Is the dropped content reconstructable from a durable artifact (git, DB, files)?
├─ YES → prefer EVICTION + POINTER — macro-compaction as a POINTER OVERLAY. Keep a
│ path/ID/summary in place of the raw bytes; the original stays reachable,
│ zoom-back intact. (Cheapest, lossless-by-reference. PD's claims/notes/tuples
│ are pointers.)
└─ NO → must SUMMARIZE. Choose:
├─ Single long thread nearing the limit
│ → IN-CONTEXT COMPACTION — macro-compaction as a SUMMARY OVERLAY: summarize
│ history *ranges*, reinitiate the window with the summary + recent artifact refs.
│ Keep the old summary in view and originals underneath; never delete outright.
│ Maximize RECALL first, then trim for precision.
├─ Long task with milestones
│ → STRUCTURED NOTE-TAKING: agent writes durable notes OUTSIDE the window,
│ pulls them back on demand. (PD: `pd note`, episodic memory.)
└─ Parallel exploration
→ SUB-AGENT ISOLATION: spawn clean-window workers, each returns a
1–2K digest. Isolation IS compaction — the parent never sees the bloat.
Boundary rule — never split a pair. Any move that summarizes or truncates message
history — macro range overlays and micro tool-output truncation alike — must shift its
boundaries so a tool_use and its matching tool_result are never separated. Drop one
and keep the other and you get a hard 400 that only /clear fixes: this is the
claude-code #14173 / #40305 failure class, and it is exactly why boundary-aware ranges
exist. Make it a precondition on every compaction step, not a cleanup afterthought.
Durable-agents cross-reference: micro/macro compaction, overlays, and boundary-aware
ranges are the in-session half of the durable-agent picture. For the full industry
comparison (Cloudflare, Temporal, LangGraph, Letta) and how Port Daddy's primitives map
onto it, see docs/research/durable-agents-landscape-2026-07.md.
3. Shared digest (the swarm's compaction) — granularity
Who reads this digest, and can they zoom?
├─ The human operator (Attention Queue)
│ → headline + counts + the ONE thing needing a decision. Must deep-link to
│ the PR/diff/note. If it can't zoom, it's over-flattened — reject.
├─ An arriving / successor agent (briefing, resurrection handoff)
│ → "what happened before you got here": open claims, recent notes, unfinished
│ obligations. Bias toward POINTERS over prose (cheaper, re-fetchable, honest).
└─ The market / billing layer (L3)
→ not prose at all: a metered ledger of tokens-per-task keyed to outcomes.
4. Metering / charging token spend across a fleet (mechanism design)
Are agents under one operator (cooperative) or across operators (strategic)?
├─ Cooperative (single-player wedge)
│ → ACCOUNT, don't auction. Per-agent/per-task token ledger, budget caps,
│ loud-fail when an agent blows its cap. No incentive problem — just visibility.
└─ Strategic (multi-operator market)
→ token spend is real COGS that must be priced. Charge the externality
(an agent that floods shared context degrades everyone). Beware sybil:
cheap identities let a fleet split work to dodge caps — tie spend to a
costly, persistent identity (see reputation/continuity).
FAILURE MODES (the context-degradation cascade)
Lost-in-the-middle starvation.Symptom: agent retrieves the right document/tool but
reasons as if it didn't; accuracy depends on where in the window the fact sat.
Root cause: models use the beginning and end of context far better than the middle.
Detection: same fact, two positions, two answers. Fix: put the load-bearing fact at
an edge; shorten the window; don't bury the obligation in scrollback.
Context rot.Symptom: quality silently decays as the session grows even though nothing
was dropped. Root cause: every token spends the finite attention budget; transformer
attention is n² and stretches thin. Detection: accuracy vs. token-count curve bends down
well before the advertised limit. Fix: compact earlier and more aggressively; budget to
effective context.
Recursive-summarization collapse.Symptom: after several compaction rounds the agent
confidently asserts things that never happened. Root cause: each summary injects a little
LLM noise; recursion compounds it into cascading hallucination. Detection: claims in the
digest with no backing artifact; drift from the source on re-read. Fix: compact from the
artifacts (git, notes, DB) on each round, not from the previous summary; keep zoom-back
links so every claim is auditable.
Over-flattening (the Scott failure).Symptom: the digest reads clean and green, but the
operator can't act because the real situation isn't reachable. Root cause: summary became a
replacement, not a lens. Detection: a digest line with no deep-link to its artifact. Fix:
every digest item must zoom; ban terminal summaries that can't be drilled into.
Compaction-as-cost-blindness.Symptom: fleet bill balloons; nobody can say which agent
or task spent it. Root cause: tokens never metered per task. Detection: no token ledger
keyed to outcomes. Fix: account first (a per-task token row), price later (only in the
multi-operator market).
Schema/tool bloat.Symptom: agents pick the right tool with wrong params; high cost,
low accuracy. Root cause: 50 tools loaded into every window. Detection: >30% of context
unused per response. Fix: 3–5 always-loaded tools, discover the rest (context precision).
WORKED EXAMPLE — a 6-agent feature swarm
Task: ship a feature across 6 parallel agents under one operator, $-capped.
Budget by role. Orchestrator: 12K (plan + six 1.5K digests). Each worker: 40K clean
window, ≤4 tools. Reviewer: 8K anchored on the diff.
Compact by reconstructability. Worker tool output that's in git → evict + pointer.
Worker reasoning that isn't → structured note (pd note) so the successor can re-read it.
Build the shared digest twice, for two readers. For the operator's Attention Queue:
"6 agents, 4 landed, 1 blocked on a claim conflict (→PR #123), 1 over budget (→kill?)."
Every clause deep-links. For an arriving 7th agent: open claims + unfinished obligations
as pointers, not prose.
Meter. One token-ledger row per agent-task. Agent 5 blows its cap → loud-fail, surfaced
in the same digest. No auction (single operator) — just visibility.
Watch the cascade. Reviewer compacts from the diff, never from worker summaries, so a
worker's hallucinated claim can't propagate into the merge decision.
Novice vs expert: novice gives all 6 agents 200K windows and one flat end-of-run summary
(bankrupt + illegible + hallucinated). Expert budgets to effective context, isolates workers,
compacts from artifacts, and ships a zoomable digest plus a per-task bill.
QUALITY GATES
Each agent has an explicit per-tier token budget sized to effective context, not the advertised window.
Every compaction step has a chosen strategy (evict+pointer / in-context summary / structured note / sub-agent isolation) with a stated reason.
Recursive summaries compact from durable artifacts, not from the previous summary.
Every digest item deep-links to its source artifact (zoom-back enforced; no terminal summaries).
The digest is produced per-reader (operator vs successor agent vs billing), not one-size-fits-all.
Token spend is metered per agent-task in a ledger keyed to outcomes; cap breaches loud-fail.
Pricing/auctions appear ONLY in the multi-operator case; the single-operator case accounts, it does not charge.
Tool exposure ≤3–5 always-loaded; the rest discovered (context precision).
A position/length sanity check exists (load-bearing facts at edges; accuracy-vs-length curve known).
