Writing
Field notes from building on the record.
On context engineering, agent coordination, world models, and holding the consequential write. Mechanisms, not adjectives.
Scope, sandbox, and gate are three different controls for an agent's blast radius.
An illustrative scenario, nothing more.
Approve everything and you approve nothing: deciding what an agent change should wait for.
Thursday afternoon.
Logging an agent change after it lands is not a gate
A shared deploy runbook sits in a store that humans and agents both read before every release. The rollback procedure has changed, and an agent is told to update that one step.
A rule in the prompt is a request. A rule on the server is a control.
A practitioner set a rule for his coding agent: keep test coverage above a threshold.
A session note is not team knowledge until a human signs off on it
Then it found a second, also marked published, saying something different.
Vendor memory answers the wrong question for a team
Open the second tool you use. Not the assistant you settled things in last week, the coding agent in a different harness. Ask it what you decided about the pricing page.
Repo-Level vs Company-Level Context for AI Agents
Repo-level context holds the code. Company-level context holds the decision.
Three tools, three context files, one team decision
We already ran the good version of the fix.
Shared Memory for AI Agents: Why Your Team Needs a Two-Tier Split
Two agent sessions. Same repo, separate branches. Both edit decisions.md.
- Context engineering
What a file cannot do
A markdown rule can describe a boundary. It cannot stop an action. That gap is the whole argument.
- World models
The rebuild tax has no invoice
Every stack change, the team rebuilds the context that made the last setup work. The cost never shows up as a line item.