Skip to content
NewApproval policies for MCP tools

Blog

Notes from the control room.

FeaturedReliability9 min read

A finished run is not a successful run

Agents don't throw. They improvise, apologise and mark the ticket solved. Here is how we detect the failures that look like success.

Maya OkaforOct 2, 2026

Approvals · 7 min read

Where to put a human in the loop

Approvals everywhere kill the point of an agent. Approvals nowhere is how you end up in the news. A framework for drawing the line.

Priya NatarajanSep 25, 2026

Cost · 6 min read

The 200k-token retry

How one retry loop turned a $0.05 run into a $4 run, and the three guards that would have caught it.

Daniel ReyesSep 18, 2026

Evals · 8 min read

Evals from production traces

Synthetic test sets miss the weird stuff. Real runs don't. How to turn yesterday's traffic into tomorrow's regression suite.

Sam WhitfieldSep 11, 2026

Security · 10 min read

MCP servers you don't control

Your agent can call tools on servers you've never audited. Allowlists, argument policies and what to log.

Lena FischerSep 4, 2026

Cost · 5 min read

Budgets that actually stop

Soft limits are a notification. Hard limits are a product decision. How we designed pause-at-100% without breaking runs mid-flight.

Tom LindqvistAug 28, 2026

Traces · 6 min read

Reading a trace waterfall

Gaps, staircases, fan-outs and long tails: the six shapes that tell you what an agent was doing before you read a single span.

Aiko TanakaAug 21, 2026