Skip to content

Latest commit

 

History

History
79 lines (54 loc) · 4.44 KB

File metadata and controls

79 lines (54 loc) · 4.44 KB

Production Results

Measured data from a 24/7 agent system running Claude Opus on a 1M token context window. 95 interactions/day via messaging bridge, 48 scheduled duty-cycle sessions. Static context re-injected every session from a portfolio of PRD and identity files.

These are not projections from a benchmark. They are measured from a running system at production load.


Per-File Compression Results

File type Original Compressed Reduction
Behavioral role specification 16,652 tokens 9,912 tokens 40.5%
Dashboard product spec 2,497 tokens 2,069 tokens 17.2%
Framework integration spec 10,977 tokens 9,879 tokens 10.0%
Infrastructure inventory 2,970 tokens 2,736 tokens 7.9%
Product specification 4,827 tokens 4,628 tokens 4.1%
Identity preload 2,876 tokens 2,623 tokens 8.8%

Why the range is wide: Compression rate correlates directly with behavioral content density. The behavioral role specification (40.5%) is predominantly narrative prose with minimal hard constraints. The framework integration spec (10.0%) is 60–65% verbatim — conditional logic, enforcement gates, and exact configuration values that cannot be compressed. The infrastructure inventory (7.9%) is 85% verbatim. The taxonomy correctly identifies what's compressible; the technique only operates there.

Files under ~2,000 tokens are excluded entirely. The fidelity risk exceeds the savings.


Session-Level Impact

Metric Value
Static context before PC 19,528 tokens/session
Static context after PC 12,535 tokens/session
Session reduction 35.8%
Tokens saved per turn 253
Savings over a 100-turn session 25,300 tokens
Equivalent additional turns freed ~12 full-complexity turns

A portfolio session that previously injected 19,528 tokens of static context now injects 12,535. The preload savings compound: 253 tokens saved per turn, every turn.


Daily Operational Savings

Source Daily savings
Bridge preload (95 turns × 253 tokens) ~24,000 tokens
PRD injections on topic-shift events (~10–15/day × ~6,739 avg saved) ~67,000–101,000 tokens
Duty-cycle preload (48 sessions × 253 tokens) ~12,000 tokens
Conservative daily total ~103,000 tokens
Peak day estimate ~175,000+ tokens
Heavy week projection 1,050,000+ tokens

Cross-System Estimates

The following estimates are derived from documented architecture specifics and community-reported overhead data for each system. They are not measured — they are modeled using the same compressibility taxonomy applied to production.

System Per-session bootstrap With PC Est. reduction Source basis
Hegematon System 19,528 tokens 12,535 tokens 35.8% (measured) Production deployment
OpenClaw active config ~15,400 tokens ~12,900 tokens ~16.2% docs.openclaw.ai + issue #21999
Hermes Agent (Telegram) ~13,935 tokens ~11,300 tokens ~18.9% NousResearch/hermes-agent issue #4379

OpenClaw context: The standard 7-file bootstrap (SOUL.md, AGENTS.md, USER.md, IDENTITY.md, TOOLS.md, MEMORY.md, HEARTBEAT.md) injects 12–20K tokens per session in an active deployment. The compressible layer — SOUL.md, USER.md, behavioral sections of AGENTS.md — is where PC operates.

Hermes context: A community monitoring dashboard found 73% of every Hermes API call is fixed overhead (~13,935 tokens). Of that, ~7K tokens is compressible behavioral system prompt content; ~6.9K is verbatim tool definitions. PC targets the behavioral layer only.


Methodology Notes

  • Token counts are approximate using GPT-family tokenizer averages (~1.33 words/token for English prose).
  • Fidelity validation is behavioral, not textual: equivalent queries run against both original and compressed variants, output compared.
  • Cross-system estimates assume the same content-type distribution observed in the measured system — actual results will vary by bootstrap composition.
  • "Conservative daily total" excludes peak topic-shift events and uses the lower bound of the PRD injection estimate.

Production deployment: Hegematon System (Hegemon Dev Framework + Automaton Agent Plugin) — github.com/swbratcher/hegemon