Seven guards that made our agents boring enough to trust
arthaudm · reddit · 2026-09-22
The author argues that most prompt-level fixes for agent failures didn't survive — the durable fixes became deterministic code around the model. He lists seven guards:
- Minimum time/evidence before irreversible tools: a voice agent can't end a call before the caller speaks; a payment agent can't submit without the final total.
- Proposal != execution: the model proposes tool calls, deterministic code checks amount, recipient, permissions, and state.
- One owner for mutable state: many agents may read, only one system writes the status field.
- Receipts after every write: read back the object to confirm it's actually visible, don't trust a 200.
- Idempotency before retries: check whether the first write landed after a timeout; blind retries cause duplicate emails and orders.
- Expiry on temporary context: context like "blocked on x" needs a close condition or TTL, or stale state leaks into future runs.
- Human approval at the real boundary: review exact recipient and words together; separate code authors from production approvers.
None of this makes the model smarter — it makes model mistakes cheaper and visible.
More from coding & agent
- Grounded Document Agent: cited PDF Q&A with LlamaParse, LlamaIndex and local Ollama — Roger_M_Taylor · 2026-09-22
- Synara v0.9.0 brings computer use to native macOS apps in beta — CurieuxExplorer · 2026-09-22
- scikit-learn co-founder, Bain AI lead to debate what agentic data science actually works — hugobowne · 2026-09-22
- Pi community ships 5 agent-team plugins as official sub-agents stay absent; Pi 0.87 splits session from context — solyarisoftware · 2026-09-22
- Pi v0.87.0 ships canonical session context editing, plus five breaking changes — solyarisoftware · 2026-09-22
- MiniMax details how to build a testbed for coding agent harness changes — MiniMax_AI · 2026-09-22