A quieter week by volume than the run leading into it, and still one with a real shift in it: the coding agent stopped being something you open a tool to reach and became something already sitting in the rooms you talk in.
The other throughline is usage limits, from two directions at once — GitHub keeps adding precision to how agent-triggered work gets billed and reviewed, and OpenAI spent the week demonstrating what happens when a limit disappears and has to come back.
21 August: the agent moved into chat
GitHub shipped agentic Copilot in both Slack and Microsoft Teams, in public preview for Business and Enterprise. Mention @GitHub in either one and the same cloud agent that runs from the CLI starts working — visibly, to everyone in the channel or meeting, rather than reporting back to one person afterward.
The capability is not new. The entry point and the audience are: a task can now start from a chat mention typed by anyone in a channel, or from an action item handed off mid-meeting, and it keeps running asynchronously afterward. The approval gate that already existed for agent-created pull requests is what stops a casual mention from becoming a casual merge — and it is worth confirming it is actually configured rather than merely available.
The same day: smaller, still real
GitHub also shipped better tools for managing blocked users — search, filtering by reason, private notes, and a cleaner separation between reviewing existing blocks and creating new ones. Unglamorous, and it fixes a workflow that gets unmanageable at scale, the same pressure behind the organization-level pull request limits from earlier this month.
20 August: a dismissal reason for risk you control instead of ignore
Code scanning gained "Mitigated" as a dismissal reason distinct from "Won't fix" — for vulnerabilities that remain in code but are controlled through something external, like a WAF rule or a network policy. It only means something if teams attach a named control, an owner and a review date to every use of it; without those three things it is just a better-sounding version of the label it was meant to replace.
25 August: OpenAI put the cap back
ChatGPT Plus Codex and Work users had their five-hour usage window lifted for a stretch this month, to mark crossing another million active users. It came back on 25 August, and OpenAI's engineering lead was specific about why: smoothing compute load, and — the more interesting admission — casual users accidentally burning a week's usage without realising it once the cap was gone. Pro subscribers stay exempt for now.
That is the same underlying tension every usage-based AI product is negotiating this month, just told from the vendor side for once: a generous-looking removal of a limit and a genuinely better user experience are not automatically the same change.
What connects the week
Every item this week is about matching effort, trust or spend to the actual shape of the work, rather than applying one setting to everything. A dismissal reason that separates "not a risk" from "a risk we control" is that. A usage window that resets rather than accumulates monthly is that, done badly and then corrected. A shared agent session in Slack is a case where the shape got more visible without necessarily getting more accountable — the open question from earlier this week about who reviews a group-watched task is still open.
What to actually do
A short list weighted toward what has a real deadline or configuration step attached.
- If you enable Copilot in Slack or Teams, configure the required-approval setting before rollout, not after the first unattended merge
- Reclassify your existing "Won't fix" backlog now that "Mitigated" exists — some of those dismissals were always mitigations wearing the wrong label
- If you or your team is on ChatGPT Plus, plan around a 5-hour window starting today rather than the uncapped access from earlier this month
- Check whether your organisation's blocked-user list has private notes on the entries that matter — the new search makes stale, contextless blocks easy to find and fix