Agent costs leak in the architecture, not the model tier: 4 patterns

Key-Register-1969 · reddit · 2026-10-24

The author argues per-run agent cost is mostly structural, not model-tier. Four leak patterns:

Unbounded retries: a stuck step replays the whole trajectory on each retry; one flaky step can burn tens of thousands of tokens for nothing. Cap retries and replay only the last error plus current state.
Tool-result bloat: full API responses, directory listings, and stack traces in the loop price every subsequent call on accumulated junk. Truncate at the tool-result layer.
Fan-out multiplication: N parallel sub-agents each get a full context copy — the bill is context × N. Keep shared state centrally and hand sub-agents a minimal brief.
No per-task accounting: teams know per-million-token prices but not per-task cost; without per-run measurement nothing gets cheaper. Log tokens per task like latency.

Takeaway: cost scales with trajectory length and context width, not output quality — design for short trajectories and narrow context.

More from coding & agent