Agents' 'retries' often hide identical tool calls — detect them structurally, not by reasoning text
ClickOk5811 · reddit · 2026-09-25
A developer observed a research agent retry a failed tool call (permission error) three times: each retry came with different reasoning narration, but the actual tool calls were nearly identical — same tool, almost same arguments, same failure.
The key insight: the narration was changing while the behavior wasn't. Reading only the reasoning trace makes it look like the agent is exploring different approaches when it's spinning in place.
His fix: a simple check that compares recent tool calls structurally instead of parsing reasoning — if the agent keeps making near-identical calls after the same failure, force it to either try a genuinely different tool/approach or stop. In theory the model should recognize this itself, but in practice you can't rely on that.
More from coding & agent
- Devin launches a redesigned home page for its AI coding agent — DevinAI · 2026-09-25
- Blender agent skill turns concept art into rigged game assets — one user shipped a Mario 64-style game in 8 hours — majidmanzarpour · 2026-09-25
- Open Manager for ComfyUI adds desktop mode for Vast.ai and other cloud GPU services — WASasquatch · 2026-09-25
- Block joins x402 Foundation, contributes Lightning payments for agentic commerce — kleffew94 · 2026-09-25
- Stripe engineer answers 'what's it like at Stripe' with demos over memos culture — schwentker · 2026-09-25
- Wasmer runs a real PostgreSQL 18.4 server on iOS and in the browser via WebAssembly — jedisct1 · 2026-09-25