Why We Ripped Out Coding Mode: Let the Model Pick Its Own Tools
Agent! used to guess whether you wanted coding or automation and hide tools to match. One broken Photo Booth task showed why the harness should stop second-guessing the model.
Early versions of Agent! had modes: coding, automation and standard. The idea sounded reasonable. A coding task doesn't need the Accessibility tool, and a "click this button" task doesn't need Xcode. Hide what isn't relevant and the model sees a shorter tool list, spends fewer tokens and makes fewer wrong picks.
On April 7, 2026, commit 7ea44c0d deleted the whole system. Here's the bug that did it, and the principle it left behind.
How modes worked
On the second turn of a task, the harness looked at what you'd asked for and matched it against two keyword lists. Words like build, compile, edit, fix and refactor meant coding. Words like click, button, window, photo and accessibility meant automation. Then it flipped a flag, codingModeEnabled or automationModeEnabled, and narrowed the tools the model could see to that mode's groups. The model could also switch modes itself through a mode tool.
The bug: Photo Booth in coding mode
The report was simply "accessibility is broken." The cause, in the commit message's own words: the auto-switch saw open -a Photo Booth as an unknown signal and defaulted to coding mode. Coding mode excluded the Auto group, and the Auto group is where the accessibility tool lives.
So the model was asked to take a photo, and the one tool built for pressing the shutter button had vanished mid-task. It did what a capable model does with the tools it has left: it fell back to osascript through the shell, which kept failing with "Can't get toolbar 1 of window 1."
Nothing was wrong with the model, and nothing was wrong with the accessibility tool. The harness had guessed wrong about intent and quietly taken away the right answer.
The fix that wasn't
The obvious patch was to add open -a to the automation keywords. The commit message calls that out directly: it would have been "one more band-aid in a long line." Keyword intent detection is a losing game. Every new app, phrasing and language adds another hole, and every miss fails in the most confusing way possible, with a model that suddenly can't do something it could do one turn earlier.
What replaced it: nothing
The whole mode system came out: 13 files, 141 lines deleted and 39 added. That included:
- the
codingModeEnabledandautomationModeEnabledflags and their tool-group lists - the turn-2 auto-switch in both the main task loop and tab tasks
predictToolGroups(), the function that guessed tool groups from your prompt- the tool-list narrowing for local endpoints in the Claude and OpenAI-compatible services
- the
codingandautomationaliases for sub-agents (you pass real group names instead)
Tools are now filtered only by your own toggles in Settings (ToolPreferencesService). Every tool you've enabled is available on every turn, and the model picks what it needs.
Two small details show the care in a deletion like this. The mode tool wasn't removed outright. It became a no-op that replies "mode switching has been removed", because a model with an old conversation in context might still call it, and a clear answer beats an unknown-tool error. And the system prompt revision went from 71 to 72, so any prompt saved on disk that mentioned modes gets re-synced.
The result
No more "Coding mode auto-enabled" log spam. No more tools disappearing mid-task. No more tool list flapping from one turn to the next.
The principle
This decision shaped a lot of what came after, so it's worth stating plainly:
The harness should enforce safety, not guess intent.
Agent! is strict where strictness is objective. It refuses catastrophic shell commands, edits to files the model hasn't read, and "done" without evidence. Those are facts it can check. Which tool a task needs is a judgment call, and the model is better at that call than a keyword list. When the harness overrules the model on a judgment call, it fails silently and confusingly. When it enforces a checkable rule, it fails loudly and with a reason.
If you're building an agent, it's tempting to "help" the model by trimming its choices. Measure first. A slightly longer tool list costs a few tokens. A missing tool can cost you the whole task.
