Google Antigravity Launches Independent Plan, Same Day Claude Code Discusses Shutting It Down
Google Antigravity's official account announced on September 25 that the 2.0 desktop version has added an independent planning mode, aligning with its CLI: input /plan, the agent first retreats to study the task, then generates an implementation plan for your review, which must be approved before execution; you can also request a lighter plan in natural language.
Almost simultaneously, Thariq from Claude Code's product side mentioned on September 23 that the team is considering shutting down the plan mode to free up Shift+Tab for the effort tier; the reason is that the model no longer requires a dedicated planning mode, while seeking feedback from loyal users on why it should remain. The community quickly reacted with phrases like "plan mode is dead" and "others just removed it while Google just added it." The official documentation still retains /plan and Shift+Tab for read-only planning, which can only be modified after approval.
Antigravity describes /plan as a multi-stage discovery: analyzing prompts, checking the workspace, reading affected files, and then delivering a reviewable proposal; within the same command set, there is also /grill-me for Socratic alignment and the paid tier /boost for multi-agent deep reasoning. When Google released 2.0 in May, commands like /goal, /schedule, and /browser were already available, with planning being more of a "plan when needed" approach, now transitioning to an explicit mode.
The plan mode in Claude Code essentially serves as a system reminder plus the ExitPlanMode tool: it prohibits writing until you confirm after research. Users have long complained that it intrudes, causes session freezes, and conflicts with context compression; some view it as a brake before risk migration. Effort, on the other hand, is another knob that controls how much computation to spend this round, from low to max, with the new model able to write plans itself at higher tiers and rewrite them based on results, rather than freezing plans.
In market mechanics, buyers want to see the developer's proposal before changing code, while sellers offer two sets of agent programming products. The driving force behind this event is the competition for keyboard shortcuts and default workflows, not whether the model can suddenly think. The funding flows towards tokens: the independent plan adds another round of read-only context, while high effort directly extends reasoning. The beneficiaries are those who internalize planning into the model loop and charge by tier; those under pressure are those relying on explicit modes for user security, yet are written off by competitors as outdated rituals. The same functionality is seen as scaffolding to be removed by one company and as a newly added selling point by another.
Source: Public Information
ABAB AI Insight
The plan mode has never been about model capability; it is a permission gate: read-only, write-prohibited, waiting for approval. Claude Code integrated it into the Shift+Tab loop because early models would directly modify repositories; Thariq wants to remove it because the new model can align in conversation and plan itself within effort, making the shortcut more valuable for "how much computation to spend this round." Google Antigravity takes the opposite approach: CLI first had /plan, and desktop 2.0 later added it, turning "research first, then act" into a demonstrable product sentence. The competition is not about whether they can think, but about which layer the default workflow is written in—prompt words, mode switches, or computation budgets.
The capital path is the interaction tax of the agent IDE. An independent mode adds another confirmation, increasing session rounds and context, while also enhancing the perception of "I am still in control"; removing the mode hands control over to effort and validation loops, making tokens sell more directly. Google needs a visible feature to benchmark against Kiro and Claude Code to prove that 2.0 is not just another chat window; Anthropic needs to give the hottest keys to tiers that can drive usage more effectively. Diffusion here materializes into a product calendar: one company believes scaffolding can be removed, while another thinks it has just been erected.
Analogies can be drawn to how even after the popularity of automatic transmissions, some still want manual modes, CI has shifted from "look at the diff before merging" to "merge only after tests pass," and every instance of making guided processes a selling point for professional software, only to remove the guide in the next version because users have already learned it. In terms of industry positioning, agent programming has transitioned from "must plan before running" to "planning has become the model's default behavior, should UI still charge separately for it?"
Structural judgment belongs to technological substitution. The mechanism is: when the model can complete exploration, alignment, and rewriting of plans in ordinary conversation, the dedicated plan mode shifts from capability to ritual; rituals remain a line on the feature list for laggards, while for leaders, they are a legacy occupying keyboard shortcuts. The gap is not in whether one can /plan, but in who still needs to make thinking a mode that must be activated.
ABAB News · Cognitive Laws
Scaffolding removed by leaders is being written as a selling point by laggards
Once models can plan, planning no longer warrants a dedicated shortcut
On the same day's feature calendar, the measurement is where the default workflows of the two companies stop at which layer