OpenCode, the open-source coding agent from Anomaly, shipped v2.3.0 on November 13, 2026 with subagents running in the background by default. The changelog says subagents now return immediately, continue working, and report back whenever they become idle, while you carry on in the parent session.
TL;DR
- OpenCode v2.3.0 (November 13, 2026) makes subagents background-first: several run at once while the parent session stays usable.
- The parent receives a report when a child goes idle, and again after each later round of work, per the agents documentation.
- The cost is visibility: transcripts fold into summaries by default.
- Background wins for slow, parallel, independent tasks; inline wins when you steer each step.
- Install via the v2 script, npm or Homebrew; verify the download against the release checksums.
What do OpenCode background subagents change in v2.3.0?
Before this release, waiting for each subagent was on you. In v2.3.0 the answer is fixed: per the release notes (PR #63382), a subagent hands control back to the parent immediately and keeps working. The parent is never blocked: it receives a report when the child goes idle, and again after each later round of work, as the agents documentation describes.
Each subagent runs with fresh context in its own child session, which is also where the no-shared-context cost comes from.
The practical effect is that you can ask the general subagent to search a large repository, start a second one on a failing test suite, and continue editing a third file yourself.
There is one safety detail worth knowing. If a subagent goes idle while its own background work is still running, its report now says it is still running (PR #63675). The parent therefore waits for the final report rather than treating a partial answer as finished. Without that, a parent could start assembling results from half-complete work.
How do background subagents compare with working inline?
The comparison below is between the v2.3.0 default and the alternative many people will want for specific tasks: staying in the loop, either by raising verbosity or by doing the work directly in the parent session.
| Factor | Background subagents (default) | Inline, watching every step |
|---|---|---|
| Parallelism | Several subagents run concurrently | One thread of attention at a time |
| Parent session | Stays free for your own work | Occupied while you watch |
| Token noise in the parent | Reports only, not every child step | Full detail in view |
| Visibility | Summaries by default | Complete, at medium or high verbosity |
| Risk | Parent may act early on partial work, mitigated by the still-running report | Slower, but easier to correct mid-flight |
| Best for | Test loops, deep searches, multi-file refactors | Serial debugging, unfamiliar code |
One trade-off matters as much as parallelism: subagents do not share the parent's context. Each child session starts fresh, so it may re-read files the parent has already read and repeat expensive model calls — and in the background, that duplicated cost now runs in parallel rather than queuing. Convenient for slow work; wasteful when the answer already sits in your conversation. Reserve the parallel default for genuinely independent, multi-step jobs, and ask the parent directly for the rest.
The pick follows the shape of the task: independent and slow work saves wall-clock time in parallel, while serial debugging is safer inline. The same trade-off applies to other agents; our look at Claude Code subagent status updates covers how a competing tool handles reporting from child agents.
What do you lose with quieter transcripts?
The same release adds "Quieter Transcripts by Default" (PR #63207). The terminal UI now folds each run of tool calls and thoughts into a single collapsible activity row. You get a readable log, but you no longer see every step unless you ask for it.
To restore the detail, set Verbosity to medium or high. An existing session.verbosity setting is kept, so anyone who already tuned it will not see a change on upgrade.
When several children run at once and you only see summaries, a subagent heading in the wrong direction is easier to miss: leave verbosity low for routine work, then raise it for the first run of a new kind of delegated task.
How do you install and verify OpenCode 2.3?
Per the OpenCode llms.txt, the v2 install script is:
curl -fsSL https://opencode.ai/v2/install | bash
It installs to $HOME/.opencode/bin - add that to your PATH. Alternatives: npm i -g @opencode/cli, or Homebrew with anomalyco/tap/opencode-v2, the recommended formula for v2. A desktop beta ships from opencode.ai/download.
Prefer verifying the binary yourself? The GitHub release page lists 30 downloadable assets with SHA256 checksums (darwin-arm64, linux-x64, windows-x64, desktop builds) plus a release attestation; the tag points at commit a6b15d3. Download the asset and the SHA256SUMS file, then run:
sha256sum -c SHA256SUMS --ignore-missing
Piping a script to bash asks for trust: read the script first, or use a package manager, on a work machine.
Which tasks suit background subagents?
The project README frames the built-in agents: build is the default with full access, plan is read-only and asks before bash, and the general subagent handles searches and multistep tasks.
Good background candidates: a test loop rerunning a suite until it passes, a deep search for every caller of a deprecated function, a multi-file refactor with few conflict risks. Poor candidates: serial debugging where each hypothesis depends on the last result, or edits touching shared state. Concurrency helps only when tasks do not collide, and OpenCode does not remove that responsibility from you.
For a wider view of the field, see our ranking of AI coding agents for 2026, the fork-versus-original take in OpenClaude vs Claude Code, and Claude Code 2.1.323 code mode vs standard tool calls.
What else shipped in OpenCode 2.3.0?
The changelog lists several smaller changes:
- Experimental MCP events behind an
mcp_eventsflag (PRs #63254, #63261, #63269). - Shell commands that survive server handoffs.
- A desktop Recent sessions sidebar showing the last 7 days, capped at 30 rows.
- A
session.forkhook for extensions. - Local model discovery for Ollama, LM Studio and vLLM waits up to 15 seconds per request instead of 1 (PR #63700), helping slow-starting local servers.
- Providers can present client certificates through
settings.tls(PR #61832). - Desktop: "Jump to an Earlier Prompt" from the palette or Cmd+Alt+G (Ctrl+Alt+G on Windows/Linux), per PR #63585.
Editor-side agents are moving in a similar direction; VS Code 1.146 and its inbox mode is one example worth comparing.
Is OpenCode established enough to depend on?
The GitHub repository shows, at the time of writing, 226,470 stars, 30,657 forks, 896 releases and an MIT licence, on a repo created April 30, 2025 (default branch dev). That is an active project, though the fast release cadence means defaults like this can change quickly.
We captured live top-10 organic results for 7 target queries to see who actually occupies the page rather than who claims to. (n=7, measured 2026-11-14) Tool-news queries are mostly occupied by fast-moving developer blogs and GitHub pages, so this article focuses on setup and trade-offs rather than repeating the release note. Third-party write-ups such as bisque.today and ai-tldr.dev cover the same release.
FAQ
Q: Can I turn off background subagents in OpenCode 2.3?
A: The docs say subagents always run in the background, and neither the changelog nor the docs describe an opt-out; raise session.verbosity to medium or high, or work in the parent session, to stay in the loop.
Q: How does the parent know when a subagent has finished?
A: The parent receives a report when the child goes idle, and again after each later round; if the child still has background work running, the report says so and the parent waits for the final one.
Q: Why do I see fewer tool calls in the terminal after upgrading?
A: Version 2.3.0 collapses each run of tool calls and thoughts into one activity row. Set Verbosity to medium or high to see every step again. An existing session.verbosity value is preserved.
Q: Is OpenCode free and open source?
A: Yes - the source is MIT licensed on GitHub. You still pay whichever model provider you connect; local models via Ollama, LM Studio or vLLM are supported for discovery.
Top comments (0)