Oliver |

I build AI systems for non-technical founders. Automations, CRMs, knowledge bases. Tested in my own companies first.

[Oliver | @olihels]

We wrote up the full checklist and where the risk actually sits: https://t.co/WWgcTo8s5q

[Oliver | @olihels]

Some teams are already shipping agents that rewrite their own instructions after failed runs, then test the rewrite before keeping it. That kind of self-correction is where this is heading.

[Oliver | @olihels]

What we're building into every client agent now: a permission layer, a memory layer that persists between sessions, and a verification step before any agent output goes live. Prompting alone covers none of that.

[Oliver | @olihels]

Model Context Protocol is becoming the standard way agents connect to your tools: calendar, CRM, inbox, database. If your stack does not speak MCP yet, that is the first gap to close this quarter.

[Oliver | @olihels]

Everyone is still arguing about the perfect prompt. The real gap in most AI agents we build is not the prompt. It is the harness around it.

[Oliver | @olihels]

Anthropic's computer use lets a model click, type, and navigate a real screen. That is powerful and dangerous. Every agent that can act needs an approval gate before it touches your calendar, inbox, or bank feed.

[Oliver | @olihels]

Most AI agent failures we've fixed this year had nothing to do with the prompt. The model was fine. What was missing was a permission gate: who signs off before it sends the email or moves the money. Build that first.

[Oliver | @olihels]

HeyGen open sourced a framework that turns HTML and CSS into deterministic MP4 video. Same input, same output, every time. That's the difference between video you can systemize and video you have to babysit.

[Oliver | @olihels]

Most teams that upgraded to Opus 5.5 never touched the effort setting. That leftover default is quietly costing tokens on every agent call. Test low and medium before assuming high is still the safe choice.

[Oliver | @olihels]

Full breakdown of what we changed and why: https://t.co/ceionb0skx

[Oliver | @olihels]

We rebuilt our whole Claude Code setup this week after realizing half our agent defaults were wrong for Opus 5.5. Here is what actually changed and what we did about it.

[Oliver | @olihels]

Dense images like floor plans and technical diagrams: let the agent crop and zoom with PIL or OpenCV before it analyzes the whole file. Tighter focus, lower token spend.

[Oliver | @olihels]

Most teams left their agents locked on max effort because that was the safe setting under the last model. With Opus 5.5 that default is now the expensive one. Medium is the balance point Anthropic recommends. Check yours.