AI Tokenomics in a Minute
Context Growth
Why agent sessions get more expensive over time as history and tool output accumulate.
In one sentence
Context growth is the increase in tokens carried across turns as prompts, files, and tool results accumulate.
Why it matters
Sessions that start cheap become expensive without anyone changing the task. Growth also raises truncation and confusion risk.
How it works
- Each turn may resend prior context unless the product caches, compacts, or resets.
- Tool outputs (logs, search dumps, large files) are common growth drivers.
- Retries multiply growth when the same bulk context is replayed.
- Mitigations: caching, summarization, narrower tools, fresh sessions for new tasks.
Example
Turn 1 sends 20k tokens. Turn 12 still carries those 20k plus 150k of command output. The “same quick question” is no longer cheap.
What this proves
Per-turn token traces prove when and how context size increased during a session.
What this does not prove
Growth alone does not prove the agent is making progress. Shrinking context does not prove correctness improved.
Last reviewed 2026-09-06. Title for citation: AI Tokenomics in a Minute: Context Growth.