OpenCode splits its coding agent in two: a read-only plan mode scouts first, a full-access build mode edits
Synopsis
The open-source AI coding agent OpenCode ships two built-in agents: build, the default with full development access, and plan, which denies file edits by default and asks permission before running bash commands, suited to exploring unfamiliar codebases or planning changes. A general subagent handles complex searches and multistep tasks and can be invoked with @general in a message. Installation covers a shell script, npm, Homebrew, pacman, mise and Nix, plus a desktop app download.
Interpretation
OpenCode divides its built-in agents into two switchable roles, build and plan, toggled with the Tab key. Coding agents have typically offered a single default working mode, leaving permission boundaries to the user. The default build agent has full access for development work, while plan is a read-only agent for analysis and code exploration.
The plan agent's read-only character rests on two concrete rules rather than prompt-level guidance. The read-only boundary is written into permission settings instead of relying on model restraint. Plan denies file edits by default and asks permission before running bash commands.
Beyond the two main agents, OpenCode includes a general subagent for complex searches and multistep tasks. Complex retrieval and multistep work need not run inside the main agent's context. The subagent is used internally and can be invoked with @general in messages.
Perspective
The material addresses developers preparing to install and use OpenCode, covering the permission difference between the two built-in agents, how the subagent is invoked, and installation via script, npm, Homebrew, pacman, mise, Nix and a desktop app. The install script picks a path in the order XDG_BIN_DIR, $HOME/bin, $HOME/.opencode/bin. The material states no follow-up experiment or roadmap.
The material states only that plan denies file edits by default and asks permission before running bash, without saying how these permissions are adjusted per project. It also gives no performance data on real tasks, so how the two tiers behave in practice remains to be seen in use.