I write about product craft, designing AI features, and keeping agents safe.
What it takes to ship an LLM feature that fails safely: render on confidence, fall back to the plain form, and measure suggestion quality from the very first commit.
Related work: AI-assisted activity creation
The four constraints a system needs before an agent can act on someone's behalf, worked through in a 2,000-line reference implementation.
Related work: A2A↔MCP bridge
Why so many company strategies fall apart, and what a working one is actually made of. Two parts, with a worked example.
Related work: Uptempo Campaign Management
What it takes to ship an LLM feature that fails safely: render on confidence, fall back to the plain form, and measure suggestion quality from the very first commit.
The four constraints a system needs before an agent can act on someone's behalf: parameter binding closes the drift attack, and consent atomicity, an independent consent surface, and destination gating close the rest. About 2,000 lines of reference Python.
The single-use property has two layers. Almost every published design enforces the wrong one. About thirty lines closes the gap.
A code-anchored walkthrough of how a2a-mcp-bridge-reference realises the security core (Vault delegation) and the two carriers that move a signed approval to the human and back: single-domain MCP elicitation emission, now implemented, and multi-domain A2A across agents.
To continue with the Zipline example, while they don't publicly share their strategy (not many companies do), here are some purely fictional possible strategic statements that Zipline could have for a 2-3 year period.
Most organizations don't actually have a strategy; they have board metrics and a pile of projects. A framework for connecting the mission, the strategy, and the work.