Blog

Notes on building agents.

Field notes from the Kairon team on evals, memory, guardrails and the teams running agents in production.

How we evaluate agents before they touch production

Every Kairon agent passes three gates before it acts on real customer data. Here is the suite, the rubrics and the numbers we watch.

Mara Okafor

September 24, 2026

Read post

Latest posts

Product

5 min read

Introducing Kairon Runtime 3.0

400+ tool connectors behind one API, durable steps that survive restarts and a trace viewer rebuilt from scratch, available today in every workspace.

Memory that forgets on purpose

Why our agents keep a short working memory, a searchable long-term store and an expiry date on every fact that sits in between.

Guardrails your agents can’t talk their way around

Policy checks belong outside the prompt. A practical guide to approvals, spend limits and tool allow-lists that agents cannot argue their way around.

How Northwind cut month-end close from nine days to two

Their finance team runs a reconciliation agent on every Payly payout. Here is the setup, the evals and what they would change next time.

Tool calling at two million runs a month

Retries, idempotency keys and the three timeouts every connector needs once traffic gets real, learned at two million agent runs a month.

Your first eval suite in an afternoon

Start with twenty real tickets, a rubric and a pass bar. You can build the rest of the harness later, once your agent passes the first twenty cases.

Halcyon Health keeps a human in every refund

A regulated support team shows how approval steps kept their agent fast without letting it act alone, and why every refund still has a named approver.

Traces you can actually read

We redesigned the trace viewer around one question: why did the agent do that? Each step now shows what it knew, what it chose and what happened.

Planning is a loop, not a step

Agents that re-plan after every tool call recover from failures that one-shot planners never notice. Here is how the loop works in practice.

Mara Okafor

July 22, 2026

Ship your first agent today.

Free up to 10,000 steps a month. No credit card needed.

Get started freeGet started free Talk to sales

How we evaluate agents before they touch production

Latest posts

Introducing Kairon Runtime 3.0

Memory that forgets on purpose

Guardrails your agents can’t talk their way around

How Northwind cut month-end close from nine days to two

Tool calling at two million runs a month

Your first eval suite in an afternoon

Halcyon Health keeps a human in every refund

Traces you can actually read

Planning is a loop, not a step

The agent builder’s digest.

Ship your first agent today.

You’re subscribed.