Santiago

Overview

Santiago is a creator, commentator, and builder whose ideas show up repeatedly across AI product, agent infrastructure, and generative media discussions. In the newsletter, he is cited both for forward-looking predictions—especially around AI video becoming real-time, interactive, and personalized—and for practical launches spanning model routing, browser automation, RAG assistants, local model usage, and agent evaluation. That combination makes him notable not just as a commentator, but as an operator testing product patterns in public.

For AI Product Managers, Santiago matters because his posts consistently surface emerging product primitives before they become mainstream: multi-model routing instead of single-vendor dependence, reusable browser skills instead of brittle one-off automations, continuous learning from user corrections, and new unit economics frameworks such as "value per token dollar." He is especially relevant to PMs building agentic products, developer tools, AI infra layers, and next-generation media experiences.

Key Developments

• 2026-05-13: Warned that depending on a single LLM provider can break an app overnight and launched an API offering access to 400+ models through one key, emphasizing flexibility and supplier diversification.
• 2026-05-14: Said Cline is among the best agentic coding harnesses, citing benchmarks where it outperformed Claude Code on Opus 4.6 and 4.7.
• 2026-05-14: Introduced a newly released Computer Use agent that can build and deploy apps inside existing applications like a human operator, positioning it as a more stable and cost-effective approach than more fragile agents.
• 2026-06-02: Launched The Grid, an OpenAI-compatible API marketplace that routes requests to the cheapest model within a selected quality tier (standard, prime, max) with a one-line code change and supplier quality audits.
• 2026-06-20: Shared that he had been running gemma-4:26b locally on a Mac Studio since April to process private documents, with local inference already handling roughly 60% of his queries.
• 2026-06-25: Shared a fully open-sourced RAG assistant for navigating airline policies, including code and a walkthrough, highlighting a concrete vertical use case for retrieval-based assistants.
• 2026-06-26: Framed agent improvement across three layers—model, harness, and context—but argued that the biggest unlock is continuous learning from every user correction.
• 2026-06-29: Open-sourced a browser automation system that records user web actions, removes retries and dead ends, and abstracts workflows into reusable skills.
• 2026-06-30: Proposed the value per token dollar metric—value created divided by token cost—as a practical way to measure agent ROI; below 1 loses money, 1 breaks even, and above 1 is profitable.
• 2026-07-07: Proposed giving AI agents their own email addresses that users can CC, enabling autonomous task handling without cluttering the primary inbox.
• 2026-07-12: Predicted AI video will evolve from static clips into real-time, interactive livestream-style experiences, including highly personalized ad formats reminiscent of Minority Report.

Relevance to AI PMs

1. Model strategy and vendor resilience: Santiago repeatedly emphasizes avoiding single-provider lock-in. PMs can apply this by designing model gateways, fallback routing, and quality tiers early, rather than retrofitting them after outages or pricing shifts.

2. Agent product design and evaluation: His work highlights practical building blocks for agents: harnesses, reusable browser skills, continuous correction loops, and clear ROI metrics like value per token dollar. PMs can use these ideas to prioritize instrumentation, memory, learning loops, and measurable success criteria.

3. New interface and experience patterns: From CC-able email agents to interactive AI video, Santiago points toward product formats that feel more ambient, personalized, and action-oriented. PMs can use these signals to prototype beyond chat—into inbox workflows, live media, and embedded computer-use experiences.

Related

• The Grid, merge-gateway, internal-ai-model-gateways, 400-models: These connect to Santiago's focus on multi-model routing, cost optimization, and reducing dependence on any single AI provider.
• Cline, Claude Code, agentic-coding-toolkit, ai-agents, llm-agents: These relate to his interest in agent harnesses, coding performance, and the infrastructure needed to make agents reliable in production.
• browser-automation-system, skill-graph, computer-use-agent, agent-working-protocol, awp-coreawp-skill: These tie into his browser automation and reusable skills approach for turning human workflows into agent capabilities.
• RAG, rag-assistant, context, memory-layer, continuous-learning: These reflect his thinking on retrieval, context handling, private knowledge workflows, and learning from user corrections.
• gemma-426b, Mac Studio, open-source: These connect to his use of local open models for privacy-sensitive workloads and practical deployment tradeoffs.
• PixVerse, pixverse-ai-v6, cinematic-realism-engine, generative-interfaces, UI: These are adjacent to his prediction that AI video will become interactive, real-time, and personalized rather than remaining a static generation medium.
• value-per-token-dollar, matrix_build: These link directly to his proposed agent economics metric and its early adoption framing.