Desktop app gets slow while agents work: decorative animations drive the CPU, the language server grows to 2.9 GB, and long conversations keep every step on the page

The Antigravity desktop app (2.26.0, Windows 11, Ryzen 7 5800H with Radeon graphics, 32 GB) feels sluggish while an agent is working, and it gets worse over a long session. I measured where the time and memory go. All numbers are from Windows process counters and Chrome DevTools metrics on the running app.

The “Working” animation

While an agent works, the conversation shows a shimmering “Working” label and a spinner. The shimmer animates -webkit-mask-position, which the browser cannot hand to the GPU compositor, so the page restyles and repaints on the main thread every frame. Over 10 minutes the window process averaged 60% of one CPU core and about 85 style recalculations per second, while JavaScript itself took under 4% of the time.

I paused the page’s animations for 16 seconds and turned them back on:

Window process CPU GPU process CPU Main thread busy Style recalcs per second
Animations running 14.6% 32.4% 21.3% 58
Paused 2.0% 21.0% 8.6% 4
Running again 18.9% 55.8% 32.8% 88

So most of the window’s work during agent runs is decoration. Typing and scrolling wait behind it. Moving the shimmer to transform or opacity, pausing it when the window is not focused, and honouring the reduced-motion setting would remove most of this.

Memory of the language server

On 25 October I sampled every process once a minute for 3 hours while one conversation was open and working:

Time language_server.exe private memory Threads Handles
16:10 326 MB 64 966
17:00 1,840 MB 79 1,147
18:00 2,643 MB 83 1,200
18:50 2,912 MB 89 1,282
19:10 2,716 MB 87 1,290

It gives memory back now and then, but each low point is higher than the one before (358, 1,690, 2,124, 2,310 MB), so what it keeps grows by roughly 12 MB a minute. On an earlier day the same process had reached 12.96 GB after about two days. I can’t tell from outside whether this is a leak or a cache without a limit. The conversation was active the whole time, so I have no idle period to compare.

Long conversations

The window’s JavaScript memory stayed flat (78 to 106 MB), so the page code is not leaking. But the number of elements on the page doubled over the 3 hours, from 4,118 to 8,299, and the number of finished fade-in animations still attached to elements grew from 541 to 1,285. Every step of the conversation seems to stay on the page with its animation attached, so each restyle has more to do as the conversation gets longer. Rendering only the visible part of the conversation would keep this flat.

Background disk work

While the agent ran, ~\.gemini\antigravity\agyhub_summaries_proto.pb (759 KB) was replaced as a whole 6 times in 90 seconds, for a net change of minus 6 bytes. That is 4.5 MB written to record a few bytes. The conversations folder also had 372 leftover -wal files (210 MB) for 500 databases, some about 14 times the size of their database, and the open conversation’s WAL stayed at 33 MB during the run.

I can share the raw samples if they help.

@Iamemily2050

Seeing the exact same pattern on Windows 10 Pro (2.26.0). Across the 6 Antigravity.exe processes on my machine, the GUI alone sits at 1,056 MB combined Working Set (460 MB WS / 401 MB Private Memory on the main renderer, with >275s of CPU time burned), while language_server.exe starts around 130 MB and ratchets steadily upward over multi-hour runs.

The compounding interaction between the -webkit-mask-position shimmer and the unvirtualised DOM is what makes deep sessions so painful: because every conversation step remains mounted in the document with its completed fade-in animation attached, each main-thread CSS restyle pass has to walk an increasingly large node tree. By turn 45-50, even basic typing in the prompt box starts stuttering behind layout work.

Pausing animations in DevTools immediately drops renderer CPU back to ~2% here too. Until the transcript view is virtualised upstream, the only reliable way I have found to keep long sessions responsive is to checkpoint active task state and uncommitted diffs into an external markdown file every ~45 turns and start a fresh conversation. Dropping the old session flushes the accumulated transcript DOM and resets the language server working set so typing stays instant.

Hi @Iamemily2050 @dllhell,

Thank you both for the excellent measurements. They’re exactly what’s needed here. Two things you can do now:

  • Reduce animation cost: Antigravity honors the operating system’s reduced-motion setting, which turns off the “Working” shimmer. On Windows, open Settings → Accessibility → Visual effects and turn Animation effects off, then restart Antigravity.
    Update to Antigravity App 2.28.0, which includes performance improvements for opening and working in long conversations.

The language server memory growth, the repeated full rewrites of the summaries file and the leftover WAL files are not addressed by these steps. If memory still grows on 2.28.0, please reply with a few hourly memory samples from that version.

Same memory growth here with long conversations — and on 2.29.0 it goes one step further and crashes the language server.

A few data points from my machine (Windows 11 Pro, 64 GB RAM) that match what you describe:

  • One active conversation had 80,006 steps: ~1.2 GB on disk, ~460 MB of actual data (~60% of the file was internal fragmentation; a SQLite VACUUM brought it from 1226 MB to 504 MB with identical row counts and integrity_check = ok).
  • At crash time the Go heap was at ~2.1 GB, and the crash is in the protobuf decoder (google_arena_unsafe) — it receives a type pointer as the allocation size, giving 0xC0000005 / PC=0x0 or a bogus “out of memory” of ~140 TB.
  • The goroutine that crashes is often the summaries reconciler, the same subsystem behind the agyhub_summaries_proto.pb rewrites you measured. Moving conversation_summaries.db and agyhub_summaries_proto.pb aside only delays the crash: they get regenerated from the large conversations and it comes back.
  • I also see oversized WAL files next to some conversations (e.g. a 15 MB -wal for a 12.8 MB database).
  • Remote Control sessions (following a conversation from the phone) make it much more frequent: crash ~1 min after the phone connects, vs ~50 min without it after compaction.

2.28.2 was stable for me (2 crashes in ~6 days); 2.29.0 crashed 33 times in under a day. Full stack signatures and a working workaround (downgrade + blocking the auto-updater): my bug report

(post deleted by author)