Google DeepMind · Antigravity
Migrating the past. Shipping the present. Governing the autonomous future.
Google DeepMind · Antigravity
Christopher Brox.
Building autonomous AI agents and self-shipping software on the Antigravity team at Google DeepMind.
00 / Thesis
Writing code is no longer the bottleneck.
Orchestrating agents is.
Agents used to autocomplete lines. Now you orchestrate them to migrate monoliths, ship features, and run the night shift.
00 / Three Eras
Where we have been. Where we are. Where we are going.
01 / The Past
Migrating what you inherited.
How real teams are clearing multi-quarter migrations in days—and the traps that break naive agent runs.
02 / The Present
Speed vs. building the right thing.
Coding agents let you build anything in an afternoon—which means you can ship the wrong thing 10× faster.
03 / The Future
Managing agents instead of writing code.
Complete automation. Engineers set intent, define policy guardrails, and review exceptions in the morning.
01 / The Past · Legacy Migrations
Every CTO in this room has a system their team is afraid to touch.
When teams first get an agent, they hand it the migration guide and type “port this repository.” It fails every time.
Here is how a real agent migration actually works—step by step, gotcha by gotcha.
01 / The Past · The Playbook
Five moves. In this order.
01
Context
Feed it the real history
02
Collaborate
Grill. Plan. Zero edits.
03
Lock
Freeze the tests
04
Slice
Pilot, /learn, fan out
05
Verify
A second agent checks
ContextCollaborateLockSliceVerify
Your wiki is lying. Git isn’t.
Give the agent every source: git history, prior PRs, production code, and the wiki.
Git History2,431 commits
Prior Pull Requests18 merged PRs
Production Code49 live endpoints
Team Wiki2022 migration doc
Antigravity
4 sources indexed
What the wiki says (3 rules wrong)What production actually does
Pay European invoices in 2 days→Paid next-day(changed 2 years ago)
Standard cent rounding→Banker’s rounding on FX(required by Finance)
Retry failed payouts 3 times→Retries 5 times(before paging support)
ContextCollaborateLockSliceVerify
Don’t prompt it. Let it grill you.
The agent asks. You answer. Ambiguity dies before code.
Antigravity·payments-migration/grill-me
/grill-me
Q1Your wiki says European payouts wait 2 days, but the code pays them next-day. Which one do customers expect?
Next-day. The wiki is three years old.
Q2There’s a custom currency rounding rule from 2019. Is that a bug to fix, or does Finance depend on it?
Don’t touch it. Month-end books won’t close without it.
Q3When a customer’s bank is down, should we retry 3 times like the doc says, or 5 times like production?
Keep 5 retries.
3 doc mistakes caught. 1 load-bearing quirk kept. Ready to plan.
Ask anything, @ to mention, / for commands/grill-me
ContextCollaborateLockSliceVerify
Agree on the plan. Zero edits.
Review one page of plan, not ten thousand lines of diff.
Antigravity·payments-migration/plan
/plan
MIGRATION_PLAN.md
Blast radius 3 services · 14 modules · 212 files
01Lock 500 golden-master test cases
02Pilot one slice end to end: refunds
03Run /learn to turn the pilot into a skill
04Fan out 7 slices · one fresh agent each
05Independent reviewer agent on every PR
Code edits: 0Approve planApproved ✓
Ask anything, @ to mention, / for commands/plan
01 / The Past · In Production · Payments Gateway
“For a payments gateway migration spanning 49 RPC methods, we pointed Antigravity at our migration guide, code generators, and 18 prior pull requests. It first flagged 10 undocumented edge cases in the guide, and then helped us run a repeatable workflow that produced 47 tested pull requests across three microservices over two days.”
antigravity.google/use-cases · Software Engineer, Payments & Commerce Infrastructure
ContextCollaborateLockSliceVerify
Agents grade their own homework. Lock the tests.
Golden master first. The agent can’t touch it.
Antigravity·terminaltests/ read-only
$ make test
✗ 2 / 500 golden-master cases failing
agent ▸ edit tests/golden_master/fx_cases.py
DENIED · tests/ is locked for agents
agent ▸ edit src/settlement/fx.py
src/settlement/fx.py +6 −2
$ make test
✓ 500 / 500 passing
01 / The Past · In Production · Framework Port
“Our team used Antigravity to refactor our user intent models from TensorFlow to JAX. We had the agent translate the model layers alongside numerical parity tests for each block, which helped us catch shape and precision differences early in the migration.”
antigravity.google/use-cases · ML Systems Engineer
ContextCollaborateLockSliceVerify
Marathon chats collapse. Slice it.
Pilot one slice. Run /learn. Fan out.
Antigravity·agent manager/learn · migrate-slice
01 · PILOT ONE SLICE
refunds✓ PR #1
/learnCaptured pilot workflow → skill migrate-slice
02 · FAN OUT 7 FRESH AGENTS WITH SKILL
payouts✓ PR #2
disputes✓ PR #3
fx✓ PR #4
ledger✓ PR #5
invoices✓ PR #6
webhooks✓ PR #7
reports✓ PR #8
01 / The Past · In Production · Reusable Workspace Skills
“We used Antigravity to analyze common patterns in our past Go code reviews and codify them into a workspace review skill. Running that pre-review check on new pull requests catches recurring style and error-handling issues early, saving about 30 minutes of review time a day.”
antigravity.google/use-cases · Senior Software Engineer, Cloud Infrastructure
ContextCollaborateLockSliceVerify
Never trust the agent that wrote the code.
Tests → reviewer agent → recording → you.
Antigravity·PR #4 · fx slicereview
Tests
500 / 500 golden-master cases passing
Reviewer agent
reviewer · fx.py:88Rounds before currency conversion. Production rounds after ($0.03 drift).
coding agentFixed in a3f9e1. Parity re-run: 500 / 500.
Browser recording
You
Waiting for code ownerApproved · merged to main
01 / The Past · Quick Room Check
Quick pulse. Which of these have you used?
Scan · select all that apply
Select all that apply
/grill-me · Agent interviews you first0%
/plan · Agree on the plan (zero edits)0%
/learn · Save pilot as a reusable skill0%
None of these yet0%
0 responses · select all that apply
01 / The Past · Honest Scope
Code parity takes hours. Cutover still takes weeks.
Collapsed from Quarters to Hours
Reverse-engineering undocumented rulesRecon
Golden-master parity test harnessHarness
Service translation & zero-drift proofSelf-Heal
Still Takes Weeks (Not Quarters)
Historical ledger & state backfillData
Downstream ERP & bank webhook cutoverIntegrations
SOC 2 / change-advisory sign-offGovernance
Eliminate the 9-month reverse-engineering bottleneck—and turn a 14-month migration into weeks.
02 / The Present · Speed vs. Judgment
You can build anything in an afternoon.
Which means you can build the wrong thing 10× faster.
Pointing fast agents at a stale Jira backlog doesn’t create velocity. It ships unused software faster.
SignalEvidenceBuild
Tickets tell you what. Analytics tells you where. Neither tells you why.
Don’t start in Jira. Connect the agent to customer reality and test the bet before writing code.
Support TicketsWhat hurts today
Product AnalyticsWhere users drop
Benchmark DataReplay logs
UserTesting MCPWhy users abandon
Antigravity
Evidence before code
Old roadmap habitAgent-first product loop
Write a 20-page PRD→Simulate on historical logs(prove lift)
Skip research for speed→Test 5 concepts in 2 hours(20 users)
File a vague Jira ticket→Send customer evidence(build brief)
02 / The Present · In Production · Product Management
“Before filing engineering requests for a new recommendation ranking model, I used Antigravity over two days to analyze historical benchmark datasets and prototype a candidate scoring rule. Running behavioral replay simulations showed a clear precision improvement, giving us solid data to scope the engineering work.”
antigravity.google/use-cases · Product Manager, Core Ranking
SignalEvidenceBuild
From customer evidence to build brief.
Connecting the agent directly to real human feedback via the UserTesting MCP.
Live Co-Presenter · UserTesting MCP
On Stage · Act II Live Walkthrough
Mike McDowell
UserTesting
01Zendesk tickets + GA4 drop-off
02Query prior video clips via MCP
03Launch 20-shopper study in 2 hours
04Winning concept → Build Brief
Press ` (backtick) anytime to switch decks
SignalEvidenceBuild
Customer words become the tests. Subagents build the rest.
Antigravity takes the UserTesting build brief, audits the data, and ships the working prototype.
Antigravity·threadline-pdp · build-briefUserTesting MCP
01 · Audit catalog dimensions first
Checked 420 SKUs · fixed 20 size mismatches (Haven 98" vs. 112") before showing scale
02 · Turn study themes into acceptance tests
Dimensions next to Buy button · auto-scaled from SKU data · zero photo upload step (18 / 18)
03 · Parallel backend & UI subagents
Shipped PDP dimension callout + live-camera View in My Room prototype + camera permission controls
04 · Close the loop with customers
Staging prototype pushed back to UserTesting MCP for verification study ✓
02 / The Present · In Production · Product Discovery to Prototype
“We used Antigravity to analyze production API usage across 400 endpoints to inform our CLI roadmap priorities, and then used it over the next few days to scaffold a multi-surface demo app connecting a mobile view, backend service, and admin console for an upcoming product review.”
antigravity.google/use-cases · Product Manager, Enterprise Cloud & Collaboration Platforms
03 / The Future · After the Roadmap
Roadmaps exist because
engineering was scarce.
Now your engineers direct swarms
that build 24/7.
engineering was scarce.
that build 24/7.
Three things that used to take a quarter now happen overnight: deal blockers, spec debates, and post-M&A integration.
DealsSpecsM&ABrief
A 4:00 PM deal blocker becomes an 8:00 AM demo.
Don’t put enterprise blockers on next quarter’s backlog. A swarm builds, tests, and records the working feature before the buyer wakes up.
Antigravity·deal-desk swarm · 6 agentsMeridian · $2.4M ARR
04:15 PM · Call transcript + CRM blocker detected
Meridian Global ($2.4M ARR) blocks Q4 close on custom Workday GL split rules
11:00 PM · Deal-Desk Swarm scaffolds tenant feature
Pulls Workday schema + contract terms → builds workday_gl_split module
03:20 AM · Adversarial verification & staging deploy
Passes 640 / 640 parity tests · Reviewer Agent ✓ · isolated customer sandbox
07:30 AM · Waiting in the Account Executive’s inbox
Recorded 45s browser demo + staging login ready for human sign-off ✓
DealsSpecsM&ABrief
Stop debating specs. Build all three overnight.
Don’t spend six weeks in PRD committees guessing which architecture will win. Spawn three swarms and let telemetry decide.
Antigravity·swarm tournament · 3 branches15 agents · overnight
BetUnblock $6.8M enterprise ERP onboarding funnel1,000 session replays
Swarm A branch/guided-wizardDiscarded
Guided 5-step modal · 48/48 tests · 92ms p99Conversion lift: +11%
Swarm B branch/one-click-oauth★ Winner · Live in 5% Canary
Zero-typing ERP OAuth sync · 512/512 tests · 68ms p99Conversion lift: +44%
Swarm C branch/chat-conciergeDiscarded
Conversational setup agent · 44/44 tests · 310ms p99Conversion lift: +19%
Overnight compute for all 3 swarms: $140Weeks of spec committees saved: 6
DealsSpecsM&ABrief
Post-M&A integration in days, not 18 months.
Point an integration swarm at both repos on Day 1. Unify identity, billing, and APIs while you sleep—so you cross-sell in week one.
Antigravity·m&a bolt-on swarm · 12 agents2 repos · Day 1 → Day 5
01 · Day-1 cross-repo cartography
Indexed 410 endpoints & 84 tables across both repos (0 docs needed)
02 · Unified identity & billing adapters
Generated tenant-isolated SAML/OIDC bridge + NetSuite & Stripe entitlement sync
03 · Locked cross-repo contract harness
Passes 1,400 / 1,400 golden-master contract tests · Reviewer Agent ✓
04 · Day-5 cross-sell live in canary
Acquired product live in primary nav (5% canary) — 18 months in 5 nights ✓
DealsSpecsM&ABrief
You can’t read 50,000 lines a day. You manage by exception.
Agents police agents inside hard guardrails. Humans only touch product taste, customer commitments, and 100% promotion.
Antigravity·morning brief · 08:00 AM33 agents · 3 swarms
Completed overnight inside guardrails (0 human prompts · < 800ms rollback)
Deal-Desk SwarmMeridian $2.4M Workday GL sandbox staged✓ 640 / 640 tests
Spec Swarm BOne-Click OAuth (+44% lift) in 5% canary✓ Reviewer + Browser
M&A Swarm410 endpoints mapped & SAML shimmed✓ 1,400 parity tests
Escalated to you at 8:00 AM (2 human decisions)
01 · Promote Swarm B (+44% lift) to 100% global?
Beat Swarm A & C overnight · 68ms p99 · 512/512 · Reviewer Agent ✓
02 · Send Meridian ($2.4M ARR) sandbox + order form?
Built from yesterday’s 4:15 PM call blocker · 45s AE demo ready to send
03 / The Future · In Production Today · Scheduled Agents
“Through scheduled tasks and CI/CD integration, we run Antigravity jobs to execute combinatorial UI test suites across our analytics platform. Each morning, engineers review the captured browser screenshots, failure summaries, and draft fixes ready for code review.”
antigravity.google/use-cases · Software Engineer, Analytics Platform
After the Roadmap · Thank You
Your codebase used to be the bottleneck.
Now the bottleneck is
knowing what’s worth building.
knowing what’s worth building.
Soon it’s how well you manage the agents
that ship it while you sleep.
that ship it while you sleep.
Connect on LinkedIn
Christopher Brox · Google DeepMind, Antigravity · ckbrox.com
Freed capacity goes to the roadmap, not the layoff list.
01 / 18