Agent reliability: provider failover, circuit breakers, and idempotent retries for tool calls
Future_AGI · reddit · 2026-09-17
A practical piece on agent infrastructure failure modes: after a write-tool call times out, you can't know if the provider rejected it or the action completed — single providers stall pipelines and blind retries create duplicate invoices. The article proposes: hot-standby provider failover behind one interface; model fallback chains (gpt-4o → Claude Sonnet → Gemini) where each model must pass the same workflow contract (tool calling, schema, context limit, latency, cost); circuit breakers with cooldown and recovery probes; and idempotent retries by persisting an operation ID before external calls. Core thesis: provider selection, breaker state, and retry policy belong in one routing layer, and the whole chain should be tested with real tool calls, not text-only health checks.
More from coding & agent
- AGNTCON + MCPCon day 1 packs the room as GitHub's MCP maintainer argues 'MCP doesn't have a context problem' — marlene_zw · 2026-09-17
- PostHog: Agent-Opened PRs Jumped From 20% to 70%, So What Do Engineers Do? — rseroter · 2026-09-17
- Non-developer builds GreekSoup, an open-source AI equity research desk, mostly by prompting Claude — Practical-Rise-1188 · 2026-09-17
- Google ships full agent lifecycle stack: context layers, self-heal, one-command deploy — blaizedsouza · 2026-09-17
- tldraw split its mascot favicon into three and serves one at random per page load — max__drake · 2026-09-17
- Stop writing LLM instructions in prose: structure criteria as JSON instead — iamrobotbear · 2026-09-17