Daily AI News — Feb 27, 2026
Three updates worth knowing about today: Microsoft is leaning hard into ‘AI that acts’ with Copilot Tasks, OpenAI is turning ChatGPT Projects into a more durable knowledge base with Sources, and the Codex CLI shipped a release that quietly improves day-to-day reliability (plus a few security hardenings).
1) Microsoft Copilot Tasks: from chat to actions
Microsoft introduced Copilot Tasks, positioning it as the shift from conversational AI to action-taking AI. The big framing: a copilot that works in the background with its own computer and browser, while you stay in control of the final decisions.
Key details (from Microsoft’s announcement):
- Research preview starts now for a small group, expanding over coming weeks (waitlist-based).
- Tasks can be recurring, scheduled, or one-off.
- It’s designed to ask for consent before meaningful actions (e.g. spending money or sending a message).
- Use cases they call out are very ‘real life’: email triage + drafts, monitoring listings (rentals/cars), briefing generation, booking appointments, subscription cleanup, etc.
Why it matters:
- This is Microsoft validating the “agent” UX pattern: tell me what you want → I’ll plan and execute → you approve.
- For builders, it’s another sign that tooling + orchestration is becoming the product, not just the model.
2) OpenAI: ChatGPT Projects adds ‘Sources’ (living context)
OpenAI’s ChatGPT release notes highlight a Projects upgrade: Sources. The pitch is simple: turn a Project into a “living knowledge base” by attaching context from where your work already is.
What’s new (per the release notes):
- Add sources by pasting links from apps like Slack or Google Drive.
- Save useful ChatGPT responses into a Project so outputs become reusable knowledge.
- Add ad-hoc text sources (notes/briefs/reference material).
Why it matters:
- This is a direct answer to the “LLMs forget everything between chats” problem: Projects are being shaped into persistent context containers.
- If you’re running teams, it’s basically pushing toward shared AI workspaces rather than individual chat threads.
Source (aggregated changelog pointing to OpenAI Help release notes): https://releasebot.io/updates/openai
3) Codex CLI 0.106.0: install scripts + realtime thread API + hardening
A new Codex CLI release (0.106.0) includes a few practical improvements for developers and teams using coding agents:
Notable highlights from the GitHub release notes:
- Direct install script for macOS/Linux shipped as a GitHub release asset.
- Expanded app-server v2 thread API with experimental thread-scoped realtime endpoints/notifications and an unsubscribe flow.
- Startup compatibility checks and a lowered validated minimum Node version to 22.22.0.
- Reliability fixes (websocket retry logic), plus a fix for a zsh-fork execution path that could drop sandbox wrappers.
- Added an explicit ~1M character input cap in the TUI/app-server to prevent hangs/crashes on oversized pastes.
Why it matters:
- The “boring” work (install, reliability, input caps) is the difference between agents you demo and agents you ship into daily workflow.
- The sandbox hardening is the kind of change you want to see show up regularly in agent CLIs.
Source: https://github.com/openai/codex/releases/tag/rust-v0.106.0
Quick take: the agent stack is becoming the UI
The common thread today is execution and persistence:
- Copilot Tasks: execution across apps with guardrails.
- ChatGPT Projects Sources: persistence of knowledge and context.
- Codex CLI: reliability + hardening for the ‘agent in the terminal’ workflow.
If you’re building products in this space: the frontier isn’t just ‘smarter models’ — it’s the boring but crucial layers that make AI dependable in real workflows.
Top comments (0)