Your agents can now keep a private journal that carries across tasks, writing and reading their own notes and having the relevant ones surface automatically at the start of a turn.
A short journal, on purpose
An agent writes an entry when something is worth carrying: a correction it had to make, how a particular person wants to be worked with, or a genuine surprise about how your systems behave. It does not write down what it did, because the task record already holds that, and it does not log a run that found nothing. Everything in the journal is read back in later sessions and spends part of the context the agent works in, so a quiet journal is the expected shape.
When an agent sharpens a lesson it already wrote, it replaces the earlier entry. The old entry stays in history, labeled with what replaced it, and stops arriving in the sessions that follow.
The agent decides what goes in its journal, and nothing approves an entry. Everyone on the account can read it.
Objectives you set, todos you can see
You can now give an agent standing objectives, see the todos it is carrying and when each is due, and have a todo with a due time run itself when that time arrives.
Objectives are yours to write and edit: what the agent is for, closed when it is genuinely met, abandoned, or replaced. Todos are the agent's own and read-only to you, ordered soonest-due first. When a due time arrives, the agent starts a task with that todo as its prompt. A run that fails leaves the item open and visible.
Due times are one-shot. Work that should repeat belongs in your handbook as a playbook with a schedule, where the whole team can see it and change it.
One block at the start of a turn, and lookups during it
Journal notes, objectives, todos, and earlier tasks on the same issue arrive as a single ranked block. Open objectives and todos go in whole; notes and earlier tasks compete for what the budget has left, relevance ahead of recency. An agent with nothing relevant to the work in front of it gets no block at all.
The block also says when it is showing only a sample, naming what did not fit and where the rest of it is, so an agent can tell "there is nothing" from "there is more than fits here". Either way it can look a subject up mid-task, by naming a person or an artifact or by asking a free-text question.
Where to find it
Open the agent and go to Work. Objectives & todos holds its objectives and the todos it is carrying, and Journal holds its entries, newest first. Both stay on the page, read-only, after an agent is offboarded.