Canon · Workspace + efficiency
Conversation-flow efficiency — eight iterations against real burn data
An analysis of where an agent's tokens actually go, measured on 152 real turns.
What the data showed
| Metric | Value |
|---|---|
| Context re-read per turn (mean, last 30) | 387,938 tokens |
| Hook injection per turn (mean) | 2,054 tokens |
| Turns with more than 1,000 tokens of hook injection | 74 of 152 (49%) |
| Output per turn, median / 90th percentile | 740 / 2,107 tokens |
Headline: half of all turns paid about 2,000 tokens for hook content the agent already had. The re-read context grows every turn, because earlier tool results and replies stay in it.
The burns, ranked
- The same hook content injected again every other turn.
- Card content echoed into the context on every create.
- Repeating a card's summary in chat after the card already has it.
- Growing a card through separate notes, which fragments it, instead of editing it in place.
- The context only ever grows.
What changed
- After writing a card, chat says the link and one line, not a re-typed summary.
- Amend cards in place. Notes are for real decision points.
- Self-heal silently, without a preface in chat.
- Card the conclusion, not the iterations. Scratch thinking isn't a durable artifact.
- Hook cadence: inject the self-check less often, and only what changed.
Each change is small. Together they compound, because every token saved in the context is saved again on every later turn.