Codex's NEW A2A Communication & Crazy NEW UPDATES: You don't know about these SECRET UPDATES!

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Codex’s NEW A2A Communication & Crazy NEW UPDATES: You don’t know about these SECRET UPDATES! · Channel: AICodeKing · 04:44 · Watch on YouTube

Topics: AI agents, AI developer tools, AI models

Summary

The video reviews Codex updates from August 22 through September 2026, including GPT-6 Astra’s experimental ability to retain and search task history. It also covers CLI task coordination with @ mentions and experimental work trees, inline questions, browser site tools, and live voice conversations. The presenter recommends trying work trees for comparing implementations and watching Astra’s searchable task history, while emphasizing that experimental features and agent-generated changes need review. Model availability, usage costs, and plan limits vary.

Key points

GPT-6 Astra’s experimental context management can preserve notes across context windows and search earlier messages and tool results within the same task.
Astra’s context feature is off by default, requires configuration and a new task, and may not be available to all users or sign-in types.
Astra costs more to use than sole, and its fast mode uses two and a half times Astra’s standard credit rate.
Codex CLI can reference other tasks with @ mentions and has experimental work tree support for separate repository checkouts.
Work trees can help isolate changes and compare implementations, but users still need to review and integrate the results.
Codex CLI can ask questions inline while continuing with work that does not depend on the answer.
Web MCP site tools let supported websites expose specific actions for Codex through the desktop app’s built-in browser.
Experimental live voice conversations are available in the terminal on supported builds, with live transcripts and microphone controls.
Codex features, model access, and usage limits vary by plan, and running several agents consumes more usage.

Chapters

Start	Chapter	Summary
00:00	GPT-6 Astra and searchable task history	The video introduces Astra’s experimental context management, its access requirements, and the trade-offs of using it.
01:31	CLI task coordination and work trees	The presenter explains how task references and separate work trees can support parallel investigation and implementation.
02:33	Inline questions and browser site tools	Codex can ask questions while continuing independent work, and supported websites can expose actions through Web MCP.
03:38	Voice conversations and closing advice	The video covers experimental terminal voice, recommends trying work trees, and notes plan and usage considerations.

Quotes

Astra can keep notes across context windows and search earlier messages and tool results from the same task. — 00:30

A work tree gives you a separate checkout of the same Git repository. — 02:03

Codex CLI can now ask questions inline while continuing to work. — 02:33

The website has to provide those tools though. — 03:07

I’d try task coordination with work trees first, especially for comparing different implementations. — 04:12

Concepts and entities

Codex, GPT-6 Astra, Codex CLI, work tree, Web MCP, Soul, Terra, ChatGPT

Full transcript (original language)

00:00 Hi. Welcome to another video. So, Codex has had a pretty interesting month. GPT-6 Astra arrived along with updates aimed at some familiar frustrations, losing earlier requirements, transferring [music] context between tasks, and waiting around for clarifications. The recent updates tackle all three. I’m covering August 22nd through September [music] 2026, and these are the changes I’d actually try in a coding workflow. So, let’s get right into it. First, Astra launched on September 3, and the feature that

00:30 interests me most is its experimental context management. [music] Normally, a long coding session eventually needs its older context compressed. The problem is that a small detail can become important much later. Maybe you said an endpoint has to remain compatible with an older mobile app, and then spent the next hour discussing something else. With this experiment, Astra can keep notes across context windows and search earlier messages and tool results [music] from the same task. So, that original requirement or an earlier test failure can be looked up again. I think this is a much more [music] useful thing to test than asking a new model to build another basic

01:00 landing page, give it a change that takes several [music] rounds, introduce a new requirement halfway through, and see whether the final result still respects the original [music] constraints. Now, this is experimental and off by default. The documentation [music] lists plus and pro users on supported clients with business, enterprise, and API key sign-in excluded at launch. You need to opt in through the configuration file and [music] start a new task, and it still needs checking. Having access to an old instruction doesn’t guarantee the model [music] will use it correctly. There’s also a cost trade-off. Astra has a higher usage cost than sole.

01:31 Its fast [music] mode uses two and a half times Astra’s standard credit rate. Personally, I try Astra on the difficult [music] debugging or refactoring work and use a cheaper model for a straightforward edit. Now, the next update is probably the one I’d recommend terminal users try first. Codex CLI gained the ability to reference other tasks with at mentions in late August. You can ask it to consult another task or send instructions to it. Then, September brought experimental work tree support directly into the CLI. A work tree gives you a separate checkout of the same Git [music] repository. That

02:03 means you can try a change without mixing its files with another task’s edits. >> [music] >> Here’s how I’d use these together. Say you’re replacing an old authentication system. Keep the investigation in one task where Codex works out how the existing login flow behaves. Then have an implementation task consult those findings and work in its own checkout. If you want to compare two approaches, give each [music] its own work tree. You can inspect both changes before choosing what to keep. Work trees already existed in the desktop app. What’s new here is the CLI support and it’s still experimental. You also need to review

02:33 and integrate the changes. Two agents finishing their tasks doesn’t tell you whether those changes work together. Another improvement makes that back and forth less annoying. Codex CLI can now ask questions inline while continuing to work. And you can answer without losing the prompt you were already drafting. For example, suppose [music] you’re adding an export feature. Codex might need to know whether you want CSV or a spreadsheet file. While you decide, it could inspect [music] the existing export code and tests. The parts that depend on your answer still need to wait. I like this because a useful [music] assistant should recognize which decisions block the task and which ones leave room to keep making progress. Now, [music] there’s a browser

03:07 update that developers should pay attention to as well. Site tools, also called Web MCP, [music] supported websites can expose specific actions for Codex to call through the desktop app’s built-in browser. Imagine a dashboard providing a tool to filter records or prepare an export. The agent [music] gets an explicit action it can use, which could reduce the amount of interface navigation needed. That gives developers another way to make their apps usable by agents. The website has to provide those tools though. This doesn’t automatically add them to every page.

03:38 The documented supported models are Soul and Terra and site tools aren’t available in enterprise or EG workspaces. So, check those requirements before trying it. Finally, the terminal got experimental live voice conversations in September. On supported builds, enable it through {slash} experimental >> [music] >> and start with {slash} voice. There are live transcripts and microphone controls. If you explain a messy bug more easily out loud, I can see [music] this being useful. I’d still type exact file names or commands when precision matters. Overall, I’d try task coordination with work trees first,

04:12 especially for comparing different implementations. Astra’s searchable task history is the experiment I’d watch most closely. I’d love to spend less time reminding a coding agent what we already agreed on. Codex is included with chat GPT plans, including free, but model access, features, and usage [music] limits vary. If you’re already paying for chat GPT, check what your plan includes before buying another coding subscription. And remember that running several agents also consumes more usage. Overall, it’s pretty cool. Anyway, let [music] me know your thoughts in the comments. If you like this video, consider donating through the super thanks option or becoming a member by

04:44 clicking the join button. Also, [music] give this video a thumbs up and subscribe to my channel. I’ll see you in the next one. Until then, bye. >> [music] [music]