Giving an Agent Read Access Without Giving It Write Access
The fix requires a read-only database role, not better instructions for the agent.
Cora Whitfield
Section
Features
13 stories in Features.
What a Stuck Agent Loop Looks Like From the Outside
Monitoring for uptime misses the expensive loops that complete successfully but waste resources.
Rafael Mbeki
Tool-Call Retries That Quietly Duplicate Side Effects
Retry logic can silently duplicate write operations when timeouts hide successful side effects.
Malik Voss
The Guardrail That Only Fires After the Damage Is Done
Moving safety checks before tool calls catches harmful actions before they cause real damage.
Estelle Kaur
Why Agent Evals Pass in CI and Fail in Production
Fixture-based evals miss production failures that only surface under live conditions.
Clara Mbeki
The 8% an Agent Shouldn't Automate
Automation works for structured data, not subjective judgment calls or missing tools.
Maya Nair
Same code, same machine, same network — the only difference was a visible window
Headless browsers trigger bot detection that headed ones easily bypass.
Theo Sinclair
A tagline match isn't a role match: what our own qualifier bug actually cost us
Automated systems can silently fail by replacing real checks with convenient proxies.
Cyrus Abara
When your own tooling blocks the thing you asked it to do — and why that's the right call
Guardrails that bend to user persuasion stop protecting you when it matters most.
Soren Bennett
Some data should never reach the model, even in an AI-native workflow
Protect sensitive data through architecture, not policy, in AI-native systems.
Rosa Fontaine
Auto-commit safety nets need a secrets-aware exception
Protecting agent checkpoints from accidentally committing secrets to version control history.
Camille Voss
The chokepoint fix: why one retry wrapper beats twenty try/catch blocks
A retry wrapper at the API boundary prevents transient failures from crashing entire batch jobs.
Mateo Haddad
The silent failure mode: when a reasoning model's thinking eats its own answer
Reasoning models can exhaust their token budget mid-thought and return nothing.
Andre Ashworth