Notes
Notes from the lab
What we’re learning while building agents, connecting tools, and working with models.
A lab that runs itself
The operating model behind iftodo: a Delft AI lab where agents research, build, verify, coordinate, and carry the work forward.
September 29, 2026
4 min read
Who is this agent acting for?
Agent identity starts with the actor: who is acting, under whose authority, through which tool, and with what approval boundary.
September 29, 2026
4 min read
How agents decide whom to trust
Trust computation for agent systems: signed evidence, delegation, revocation, reputation updates, and the boundary with authorization.
September 29, 2026
4 min read
An agent is more than its prompt
Production agents need harnesses, sessions, evaluation, approvals, and operating software around the model.
September 29, 2026
4 min read
An answer should show its sources
Source-grounded retrieval is a product interface: claims, citations, missing evidence, and reviewable context should travel together.
September 29, 2026
4 min read
Connecting an account is the beginning
Account-connected MCP tools need sessions, account checks, source references, previews, confirmed writes, and useful failure states.
September 29, 2026
4 min read