`Libretto` turns AI browser automation into inspectable code

Instead of trusting runtime prompts, browser flows are generated ahead of time as scripts you can read, version, and debug. That makes flaky portal automation more maintainable, especially when DOM-based agents keep breaking; worth testing if reliability beats pure agent runners.

[ KEY POINTS ]

Built after a year of maintaining healthcare portal automations, where brittle flows and failed runs consumed too much debugging time.
Takes a hybrid path: Playwright for UI plus in-session network/API calls. That trades fragile DOM parsing for more reliable execution and better bot-evasion.
Targets the main pain of tools like Browseruse and Stagehand: repeated AI calls, opaque runtime behavior, and weak support for debugging or updating flows.
Includes manual action recording, step-through debugging, and an optional read-only mode to avoid accidental submissions while iterating.
Generated code follows repo conventions and abstractions, so the automation lives inside the normal engineering workflow instead of a separate agent black box.

Originalgithub.com/saffron-health/librettoRead original →

// related

#0001
#0001Agents & tools Google AI Forum60 minutes ago
`Google Antigravity` shifts toward weekly quotas, hurting long coding sessions
60radar
Google AntigravityAI coding agent — developer tool powered by Gemini
Short-reset Flash access is gone, pushing heavy use into 7-day quotas. The tool now fits burst work better than daily agentic coding.
- Gemini 3.0 Flash with roughly 5-hour resets was the practical daily driver; its removal breaks predictable iteration loops.
- Paid Ultra users are also hitting tighter limits, so this is not just a free-tier downgrade.
- Weekly cooldowns turn debugging, refactoring, review, and test loops into prompt budgeting. Keep a fallback agent ready.
Source: discuss.ai.google.dev/t/google-antigravity-has-come-to-aRead original →
FIG-0011:1
60radar
FIG-0011:1
#0002
#0002Agents & tools Simon Willison4 hours ago
`Gemini Spark`, Google’s hosted agent tied to Workspace apps
60radar
Gemini SparkHosted AI agent — native Google app connections
Google is packaging app-connected agents around Workspace, but much is still coming soon. Security and credential handling decide whether it becomes useful or risky.
- Gemini Spark connects natively to Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube, and Maps. That makes it closer to a work agent than a chat UI.
- FAQ says it runs on Gemini 3.5 Flash and Antigravity. The Antigravity stack spans a desktop app, CLI, SDK, and VS Code fork.
- Enterprise notes mention fresh isolated ephemeral VMs, Agent Gateway DLP, and encrypted credentials. That is the right threat model, not proof it is solved.
- Since the product is not broadly testable yet, treat it as a roadmap signal. Do not build critical automations around it until GA behavior is clear.
Source: simonwillison.net/2026/May/20/google-io/#atom-everythingRead original →
60radar
PHOTO
FIG-0021:1
#0003
#0003Agents & tools GitHub Changelog4 hours ago
`GitHub Copilot` in VS Code gets task-based auto model routing
70radar
Model choice moves from manual picking to routing by task, utilization, and health. Useful for lower-friction coding, but less control over exact model behavior.
- GitHub Copilot now selects a model using task fit, utilization, and model health metrics, aiming for reliable and token-efficient runs.
- The change matters most in VS Code, where model switching interrupts small coding loops; default auto mode should reduce that friction.
- Tradeoff: less explicit model control. Keep manual model selection for debugging, refactors, or prompts where output variance matters.
Source: github.blog/changelog/2026-05-20-auto-model-selection-noRead original →
FIG-0031:1
70radar
FIG-0031:1

`Libretto` turns AI browser automation into inspectable code

// related

`Google Antigravity` shifts toward weekly quotas, hurting long coding sessions

`Gemini Spark`, Google’s hosted agent tied to Workspace apps

`GitHub Copilot` in VS Code gets task-based auto model routing