Topic hub
AI Agents
Build autonomous AI agents that reason, use tools, and accomplish complex tasks.
All posts · 51
AI AGENTS01
Kubernetes Agent Sandbox: Self-Hosted Setup on Your ClusterBy default, agent-sandbox v1.0.8 lets sandboxes reach any public IP on any port and points their DNS at 8.8.8.8. The on-prem settings that close it.
AI AGENTS02
OpenAI Agents SDK With Ollama: What Still Calls OpenAIAt v0.23.1 the SDK's trace exporter posts prompts and tool outputs to a hardcoded api.openai.com URL whenever a key resolves. Wire Ollama and close that path.
AI AGENTS03
LangGraph Checkpointer Postgres: PgBouncer and SSL ErrorsAt 3.1.2, the LangGraph Postgres checkpointer factory hardcodes prepare_threshold=0 and pipelines every write. Six errors traced, two undiagnosed.
AI AGENTS04
Mem0 vs Zep vs Letta vs Cognee: Which to Use in 2026Mem0 publishes 94.4% on LongMemEval, Zep 90.2% at 104ms p50 retrieval. The accuracy argument is over, so pick on tokens, latency and certs.
AI AGENTS05
AI Agent Tool Failure Recovery: Retry, Re-route or ReplanA targeted hazard hint lifts agent accuracy by 25.5 to 35.5 points. More inference budget buys 3.5 to 11.5. Tool failure recovery starts with diagnosis.
AI AGENTS06
Anthropic API Error 529 Overloaded: Agent Retry PolicyThe official SDKs retry a 529 twice, for 1.5 seconds of nominal backoff. Here is the branch table an agent loop needs, including three 429s.
AI AGENTS07
The 400 names input_json_delta because it is a streaming delta type, never a request-side content block. Six documented breaks in streamed tool_use assembly.
AI AGENTS08
Is It Safe to Open an Untrusted Repo? 12 Coding Agent DefectsTwelve fixed defects across three vendors let a repository configure the coding agent before the trust prompt. What trust gates, and what it never does.
AI AGENTS09
tool_use ids were found without tool_result blocks: 6 FixesThe 400 does not always mean a result is missing. In two filed reports every result was present, just not ahead of the other blocks in that message.