cd /news/ai-agents/agent-shell-0-98-updates · home › topics › ai-agents › article
[ARTICLE · art-173092] src=xenodium.com ↗ pub= topic=ai-agents verified=true sentiment=↑ positive

agent-shell 0.98 updates

Agent-shell 0.98.2 adds DeepSeek Harness support, bringing the Emacs-based AI agent client to 24 supported agents, and requires Antigravity users to run `agy acp` instead of a separately downloaded binary, according to the release notes from maintainer xenodium. The update also adds blockquote rendering with embedded Markdown, image thumbnails inside tables, link-target hover previews, and two new entry points, `agent-shell-markdown-open-link-at-point` and `agent-shell-markdown-xref-backend`, plus a `agent-shell-notice-filter-function` for filtering agent stderr notices. Other changes include `M-RET` on an empty prompt to steer the last queued prompt into the running turn and a `(+N)` menu for agent options that do not appear in the buffer header.

key takeaway

agent-shell 0.98.2 now runs Antigravity via `agy acp` (install a recent CLI, put `agy` on PATH, run `agy acp --login`) and adds DeepSeek Harness as its 24th supported agent, started with `M-x agent-shell-deepseek-start-agent`.

read6 min views2 publishedNov 20, 2026

Another month, another agent-shell update. If you missed the last post, have a look at the 0.87 update. This time around, the most notable changes are around rendering, performance, and bugfixes, sprinkled with new features here and there.

As usual, this post showcases highlights, but please check out the full list of changes if you're after the nitty-gritty.

agent-shell? agent-shell brings your favorite AI agents to Emacs, keeping you at home: familiar buffers and bindings, no terminal emulation. All powered by ACP, the Agent Client Protocol.

DeepSeek Harness joins the family, bringing the total number of supported agents to 24. Start DeepSeek with M-x agent-shell-deepseek-start-agent, or pick it from M-x agent-shell as usual.

agy (action required) Latest Antigravity ships with ACP out of the box, so agent-shell now runs agy acp rather than a separate binary you had to download and unpack. If you're an Antigravity user, install a recent CLI, make sure agy is on your PATH, and sign in once:

agy acp --login

Blockquotes improved their display, rendering embedded Markdown (#942).

They also joined the TAB navigation party. TAB and S-TAB now visit blockquotes, stopping on their first quoted character.

With point on a blockquote, RET now copies the quote's raw Markdown, without the > markers.

Images are now rendered as thumbnails within tables. Press RET on an image to open in full view.

Thank you @alberti42 for #938. Tables also realign in the resized window's own buffer now (#934), so splitting a window no longer leaves a table laid out for the window it used to be in.

Hovering a rendered link (or with point on one) now shows its target URL, so you can judge whether it's worth opening.

Two new entry points landed alongside (#933):

agent-shell-markdown-open-link-at-point opens the link at point.agent-shell-markdown-xref-backend plugs into xref-backend-functions, so M-. on a rendered path:line citation visits it. It returns nil anywhere else, web and binary links included, leaving point to the next backend. RET toggles the fragment at point, as usual. C-u RET now toggles groups recursively (#924, #945).

Queued a prompt, but wanted to steer instead? Pressing M-RET on an empty prompt now offers to steer the last queued prompt into the running turn (#918, #919). While this isn't something I need often, I'm pretty happy with the UX we landed on.

While on the topic of steering, chat mode now handles steered prompt labelling (#930, proposed by @Gleek). If you'd like to render labels differently, agent-shell-chat-label-function now receives the :steered param.

Agents often expose options that never make it to the buffer header. Additional options are now offered via the (+N) menu (#921).

Agents write all sorts to stderr, and agent-shell surfaces it under "Notices". You can now use agent-shell-notice-filter-function to filter notices (or modify) as preferred (#950 by @astoff). It receives an alist with :text, one stderr line, and returns the text to show or nil to drop it:

(setopt agent-shell-notice-filter-function
        (lambda (notice)
          (unless (string-prefix-p "[idle-compaction]" (map-elt notice :text))
            (map-elt notice :text))))

A filter that signals or returns a non-string is reported, and the notice shown unfiltered.

agent-shell-system-prompt is appended to the agent's default system prompt, sent whenever a session is created, loaded, resumed or forked (acp.el #37 by @systemfreund):

(setopt agent-shell-system-prompt "Keep answers short and respond like a pirate.")

Only agents advertising ACP's systemPrompt capability can take one. Configure a prompt for an agent that can't, and agent-shell says it's ignored rather than dropping it quietly. An agent config's :system-prompt takes precedence.

agent-shell-make-environment-variables now takes a function as a value and calls it when the agent starts, rather than when your config loads (#948 by @nbarrientos). Handy for a password store still locked at init:

(setopt agent-shell-opencode-environment
        (agent-shell-make-environment-variables
         "MY_SECRET" (lambda ()
                       (auth-source-pick-first-password :host "example.org"))))

A function that errors or returns nil leaves the variable out of the agent's environment and says so in a boxed warning in the shell. Works across all agents.

project-switch-project Starting a shell from C-x p p used to land in the current shell's directory rather than the project you just picked, and offer that directory's sessions to resume. It now starts in the chosen project (#949 by @OSadovy).

When agent-shell-session-fork-restore-history is non-nil, forked sessions will attempt to restore interaction history as per agent-shell-session-restore-verbosity.

agent-shell-session-fork-restore-history is off by default, so you'll need to enable it with:

(setopt agent-shell-session-fork-restore-history t)

Note: You'll need an agent that supports a session right after forking it, like claude-agent-acp 0.90.0 or newer.

Streaming performance received some love. Message chunks are coalesced before rendering (#883 by @moyotar), thought chunks are batched the same way, and restored sessions replay with far less overhead (#897 by @Gleek).

As mentioned, the much effort in the release went into fixes and improvements.

Rendering:

Performance and responsiveness:

agent-shell-cwd-function is called once per command (curl --version check that ran on every submission is gone (SIGTERM before being killed on shutdown ( Elsewhere:

flymake-mode is on ( A few more additions for package authors building on agent-shell.

agent-shell-prompt-ranges returns the range of every prompt in the buffer, in buffer order, covering prompts typed at the shell, replayed by a restored session, and steered into a running turn (#882). Each comes back as an alist spanning the comint shell prompt itself (:prompt-start, :prompt-end) and the input typed after it (:input-start, :input-end).

_meta agent-shell-subscribe-to now offers a usage-update event, broadcast when the agent reports context window or cost usage. Events now also carry ACP's _meta, so integrations can read whatever an agent attaches without waiting on agent-shell.

agent-shell-session-strategy now accepts a function. It's called with the resumeable sessions and returns new, latest, prompt, or a (session-id . ID) cons to load a specific one:

(setopt agent-shell-session-strategy
        (lambda (context)
          (if (map-elt context :sessions) 'latest 'new)))

If you peeked at the commit logs for the period, you'll see it's been another busy month. Since the last post, 193 commits landed, 32 issues have been closed and 27 pull requests merged. As of this writing, the backlog sits at 6 open issues and 5 open PRs (versus 4 and 6 last time around).

Zooming out a little, here's how the backlog has tracked:

As usual, this chart was generated using the /github-activity skill shared in my emacs-skills repo.

All of this requires daily attention.

While paying for tokens keeps agents running, sponsoring keeps agent-shell running.

agent-shell is built and maintained by me, an indie dev, while the tools it competes with often have well-funded teams behind them. Funding Emacs-native tools, so we can keep living in our beloved editor, matters now more than ever.

Using agent-shell at work? Ask your employer to sponsor too. Companies can contribute at a scale individuals can't.

Thank you to all contributors for these improvements!

agent-shell stands on acp.el to speak ACP and shell-maker for its shell foundations, so fixes landing there reach you just the same:

As of this writing, we're 89% of the way to the 150 monthly sponsors goal, a tiny closer since the last update. If agent-shell is part of your daily workflow, please consider becoming a monthly sponsor (or ask your employer to).

Not in a position to sponsor? A GitHub star helps with exposure and brings in new users.

── more in #ai-agents 4 stories · sorted by recency
── more on @agent-shell 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/agent-shell-0-98-upd…] indexed:0 read:6min 2026-11-20 · —