After Generating Several Candidates, Which One Do You Adopt? Designing Best-of-N That Arbitrates by Verification
With Gemini 3.5 Flash's speed, generating several implementations of the same task has become practical. The hard part is no longer generation but arbitration. Here is the design and TypeScript implementation of a Best-of-N arbiter that picks the winner using verifiable signals only — not majority vote, not self-reported confidence.
Accounting for Which Agent Spent What: A Cost Attribution Design by Task
Your month-end bill is one number, but running multiple agents on Gemini 3.5 Flash hides which task ate the cost. Separate from a budget guard, I share a cost-attribution accounting design that maps usage to per-task and per-site cost, with a solo-operator implementation and numbers.
Tracing Parallel Agents After the Fact: Observability with Structured Logs and Spans
Running multiple agents in parallel on the Antigravity 2.0 desktop makes it impossible to tell which one is doing what. I share an observability design that drops tangled print debugging for run_ids and spans you can trace afterward, with a solo-operator implementation and numbers.
Making Managed Agent Batches Safe to Re-run: Idempotency and Checkpoints
Running overnight batches on the Antigravity 2.0 Managed Agents API makes recovery from partial failure unavoidable. Starting from a duplicate-post incident, I share the implementation of idempotency keys, a checkpoint store, and resume logic, with real numbers from solo operations.
Supervising Multiple Agents at Once on the Antigravity 2.0 Desktop: Screen Layout and Interruption Design
How I lay out the screen, score interruption order, and measure my own ceiling on parallel agents in Antigravity 2.0 — with the small scripts and the numbers I actually recorded.
Containing Failure in Antigravity Multi-Agent Systems: Three Boundaries That Stop Cascades
Antigravity multi-agent setups run beautifully in isolation but cascade in production, where one small failure drags the whole orchestration down. These notes organize the fix around three boundaries—layered control, trust separation, and observability with idempotency—down to the TOML and the correlation-ID wrapper.
Designing Schema Evolution So Sub-Agent Handoffs Never Break
Put a typed contract at the boundary where a downstream agent receives an upstream agent's output, and learn how to evolve that schema without breaking existing flows — with validation code, a migration sequence, and the production symptoms to watch for.
Implementation Notes: Running AdMob Mediation A/B Tests with Antigravity Parallel Agents
Notes from automating AdMob mediation A/B testing using Antigravity parallel agents across an indie wallpaper-app portfolio with 50M cumulative downloads. Covers joint eCPM + ARPU optimization, the 7-day rotation pipeline, and the rollout design that does not break production.
Splitting Daily Crashlytics Triage Across Five Antigravity Sub-Agents
Running six indie iOS and Android apps, the morning Crashlytics triage was draining me. I split the workflow across five Antigravity sub-agents (Fetcher, Classifier, Repro, Patch, PR) and locked their input and output to JSON schemas. Two weeks of production data shows where the human review boundary belongs.
Enterprise Multi-Agent Skill Management with gh skill — Profiles, CI Validation, and Cross-Agent Consistency
A complete architecture for managing SKILL.md across multi-agent teams. Covers role-based skill profiles, three-layer agent compatibility design, GitHub Actions CI validation, and automated release notifications.
gh skill: Sharing AI Agent Knowledge Across Claude Code, Copilot, Cursor, and Gemini CLI
The gh skill GitHub CLI extension lets you package SKILL.md definitions and distribute them across 30+ AI coding agents. Here's how it works and how to get started.
A Practical Guide to Designing Multi-Agent Workflows with Google ADK
Learn how to design and implement multi-agent workflows using Google Agent Development Kit. Covers orchestrator and sub-agent role design, state passing, parallel execution, and graceful error handling with real examples.