engblogs

Sea's Codex-powered agentic AI strategy reframes software development at scale, turning engineers into system orchestrators who leverage AI-assisted workflows, test-driven prototyping, and CI/CD across distributed microservices.

A concise overview of Code World Models and AutoHarness as two complementary approaches for enabling robust LLM-driven agents in complex, partially observable environments.

Explores Code World Models vs AutoHarness for LLM-driven game agents: learning a full world model versus evolving a minimal code harness to prevent invalid moves, under partial observability and stochasticity, with planners like MCTS/ISMCTS.

Building AI-native, geospatial-first systems for federal infrastructure by integrating mechanistic interpretability, external memory, and grounding across orchestrated models to accelerate permitting and improve decision accuracy.

A deep-dive into building AI-native, multimodal geospatial systems for federal infrastructure, exploring mechanistic interpretability, external memory, and model orchestration to accelerate permitting while ensuring grounding and regulatory compliance at national scale.

Modernizing GitHub Issues navigation performance through a local-first client-cache model that renders from IndexedDB with stale-while-revalidate, plus preheating and a service worker to deliver instant navigation across soft, turbo, and hard navigations.

Modernizing GitHub Issues navigation from latency to instant using a local-first caching stack (IndexedDB and in-memory), stale-while-revalidate, preheating, and a service worker to accelerate hard, turbo, and soft navigations.

Scalable, multi-tenant marketing for hundreds of retailers that preserves brand isolation and delivers Marketplace-grade personalization via a per-retailer workspace, reusable Campaigns Engine, event-driven batching, and a provider-agnostic delivery layer.

Analysis of OpenAI's response to the TanStack npm supply chain attack, outlining certificate revocation, notarization controls, and macOS update timelines to safeguard users and verify software provenance.

Designing a multi-binary, elevated sandbox for Codex on Windows that balances safety and developer productivity through OS-level isolation, synthetic SIDs, write-restricted tokens, and firewall-enabled controls.

OpenAI's response to the TanStack npm supply chain attack, detailing credential rotation, code-signing certificate revocation and macOS update guidance to protect user data and software provenance.

Balancing safety and usability in a multi-layer Windows sandbox for Codex, detailing synthetic SIDs, write-restricted tokens, an elevated setup, and firewall-based network controls to enable open-ended developer workflows without requiring admin prompts.