Blog

• Notes on the design of Kestrion 5 November 2026The thinking behind Kestrion, an open-source framework for agents that resume after failures.
• Cost and latency budgets for agents 2 November 2026Set limits early, measure per run, and spend the expensive model only where it helps.
• FastAPI for agent backends 30 October 2026A simple pattern for exposing long-running agent runs over HTTP.
• Managing the context window 27 October 2026Practical ways to keep prompts small, relevant and affordable as a run gets longer.
• Structured output and schema validation 24 October 2026Validate everything the model returns, and give it a bounded chance to fix mistakes.
• Timeouts and cancellation in long-running agents 21 October 2026Every step and every run needs a deadline, and a way to stop cleanly.
• Evaluating LLM pipelines with small datasets 18 October 2026Thirty good test cases are far better than none. How to start and what to track.
• Testing agents without flaky tests 15 October 2026Fake the model for most tests, and keep live-model tests few and clearly separate.
• Logging and tracing agent runs 12 October 2026What to record so that a bad agent run can be understood after the fact.
• Approval steps before risky actions 9 October 2026How to pause an agent for a human decision without losing the run.
• Keeping agent services stateless 6 October 2026Why the run state should live outside the process, and what that buys you.
• Validating generated SQL before it runs 3 October 2026Parse the query, restrict what it can touch, and cap what it can cost.
• Pitfalls in natural-language-to-SQL 30 September 2026Why turning questions into SQL is harder than it looks, and where the errors come from.
• Designing MCP tools that models can use 27 September 2026Practical rules for naming, describing and shaping tools so that models call them correctly.
• The Model Context Protocol in plain terms 24 September 2026What MCP is, what a server and a client do, and why it is useful.
• Retries and backoff for LLM calls 21 September 2026Which errors to retry, how long to wait, and when to stop.
• Making tool calls safe to repeat 18 September 2026Idempotency keys stop a retry or a resume from doing the same thing twice.
• Checkpointing state in LangGraph 15 September 2026How LangGraph checkpointers save progress, and what to keep in the state object.
• What durable execution means for agents 12 September 2026A plain explanation of durable execution and why long-running agents need it.
• Why agents fail in production 9 September 2026Most agent failures are not about the model. They are ordinary distributed systems problems.