D6 — Chatty agent loops

The problem

Agent frameworks love small steps: observe, think, call a tool, repeat. Each step re-sends the whole conversation context. Ten 80-token answers in a burst, each carrying 1,200 tokens of re-sent context, is the signature of an agent doing one task's work at ten tasks' input cost. This is the agent-fleet waste class — the reason coding-agent bills surprise people.

Detection

Calls are grouped into runs: same route, short completions (under 300 tokens), consecutive calls within a 600-second anchor of each other. A run is flagged as an agent loop when it is long enough (n ≥ some minimum) and shows the re-read signature — a majority of the run's calls sharing one prompt prefix, proving the same context is being re-sent. Runs too short to matter are kept silent.

Savings estimate

The estimate models batching the loop's steps (default batch factor 5) and charges the re-sent context for the calls batching would eliminate, at the run's minimum observed rate — never an average that a cheap outlier could inflate:

saved_calls = n − ceil(n / 5)
overhead    = run-median prompt tokens        (the re-sent context)
savings     = saved_calls × overhead × minimum effective billed rate in the run

"Effective billed rate" prices the re-sent context as it was actually paid for: context served from provider cache counts at the cache-read rate, not the full input rate. An agent loop that already caches well still wastes calls, but the report won't claim full-price savings for tokens you bought at a tenth of the price.

Worked example (golden fixture)

12 claude-haiku-4-5 calls 65 seconds apart, 80-token completions; run clustering yields one 10-call run (span 585s) plus a 2-call run kept silent. Six of the ten calls share one prefix — re-read signature met, "agent loop suspected":

saved_calls = 10 − ceil(10/5) = 8
overhead    = 1200 tokens (run-median prompt)
savings     = 8 × 1200 × $1 / 1e6 = $0.0096   →  monthly (×10) = $0.096

Known limitations

Batching is an engineering change, not a config flag — the estimate tells you what the loop's chattiness costs, not that a 5× batch is free to implement. Runs without prefix evidence (no hashes in the export) fall back to timing heuristics and are labeled at lower confidence.