GenAI PMGarry Tan
Investor and startup leader mentioned sharing eval tooling and operational advice for agents and CI workflows. The newsletter links him to repositories and experimental modes used for agentic engineering.
Key Highlights
- Garry Tan is repeatedly cited for practical guidance on agent evals, coding workflows, and production-oriented AI harness design.
- He shared the gbrain-evals repository with 40+ categories and stressed measuring retrieval, latency, and cost per compute.
- He described using Capy.ai, GStack, and GitHub/CI workflows to accelerate bug fixing and pull-request execution.
- He argued that better orchestration and harness design can outperform raw model usage alone, even with the same frontier models.
- He also highlighted infrastructure details such as Tailscale, MCP, and real Chromium browser access as important enablers for agent products.
At a glance
- Role
- Investor / startup founder
- Company
- Y Combinator
- X / Twitter
- @garrytan
Garry Tan
Overview
Garry Tan is an investor and startup leader who appears in the newsletter as a hands-on practitioner of agentic engineering workflows. Across recent mentions, he is associated with sharing evaluation tooling, operational advice, and practical patterns for using AI agents in software development, especially around pull requests, CI workflows, browser-based automation, and model coordination.For AI Product Managers, Tan matters less as a generic commentator and more as a signal of emerging operator best practices. His posts connect concrete repos and workflows—such as `gbrain-evals`, GBrain, GStack, Capy.ai, and agent harnesses—to tactical questions AI PMs face: how to evaluate agent performance, reduce wasted token spend, coordinate multi-agent work, and build production-ready development loops around frontier models.
Key Developments
- 2026-09-18: Garry Tan said every AI harness should support Tailscale, noting that Muse and Grok Bot support it out of the box while Codex and Claude Code cloud containers currently do not.
- 2026-09-22: He shared an example of Capy.ai executing an ambitious bug-fix wave on GBrain, highlighting task delineation, automatic parallelization, and a clean GitHub pull-request and CI workflow.
- 2026-09-24: He said he uses Capy.ai daily on all GStack and GBrain pull requests and is moving at least 4x faster than with raw Codex or Claude Code via Conductor.
- 2026-09-27: He shared a preferred bug-fixing workflow using Capy.ai with GStack `/autoplan` on a production issue and GPT-6 medium reasoning.
- 2026-09-28: He recommended the AsideAI browser with MCP for agents running into antibot issues, suggesting a spare always-on computer so agents can use real credentials through a real Chromium browser.
- 2026-10-02: He highlighted Modus’s open-source test for evaluating whether AI can perform actual junior-auditor work, and predicted that many jobs will eventually have similar benchmarks.
- 2026-10-03: He highlighted Capy’s model-agnostic cross-session coordination, describing how a GBrain collaborator used multiple Capy threads to route around existing work and start a coordinated fix wave.
- 2026-10-04: He said he deleted roughly a thousand lines of markdown from GStack because frontier models have improved enough that many older prompting tricks are no longer necessary; he also argued harnesses can coordinate the same models into better coding outcomes.
- 2026-10-06: He commented that lab-developed harnesses are incentivized to burn tokens, while startup-developed harnesses can provide utility by turning usage into deterministic, tested, repeatable code; this came in the context of Grep and agent workforce tooling.
- 2026-10-11: He shared the `gbrain-evals` repository, describing 40+ categories that can each be auto-researched and auto-improved, and said clear evaluations for retrieval, latency, and cost per compute are useful for GBrain-style data tasks. He also said Capy.ai works best with experimental modes such as Captain enabled.
Relevance to AI PMs
1. Evaluation-first agent products: Tan repeatedly points to evals as the foundation for useful agent systems. AI PMs can apply this by defining benchmark categories for real workflows—such as retrieval quality, latency, cost per compute, and task completion—before optimizing UX or model choice.2. Operational patterns for agentic software delivery: His examples around GBrain, GStack, Capy.ai, GitHub pull requests, and CI show what a production-oriented agent workflow looks like. PMs building coding agents or internal developer tools can use these patterns to scope features like autoplan, parallel tasking, PR generation, and test-aware execution.
3. Infrastructure and environment design matters: His comments on Tailscale, Chromium-based browser control, MCP, and antibot constraints underline that product performance often depends on environment access, not just model intelligence. AI PMs should treat secure connectivity, credentialed browsing, and real-world execution surfaces as core product decisions.
Related
- GBrain / GStack / gbrain-evals: Core projects repeatedly linked to Tan’s workflow, especially around evals, bug fixing, and agent-driven development.
- Capy.ai / Captain: Central to his examples of bug-fix waves, cross-session coordination, and experimental modes for better agent performance.
- Claude Code, Codex, Conductor, GPT-6 medium reasoning: Referenced as comparison points or components in his preferred coding-agent stack.
- AsideAI, MCP, Chromium: Connected to his advice for browser-based agent execution in environments with antibot friction.
- Tailscale, Muse, Grok Bot: Used in his discussion of the infrastructure support AI harnesses should provide.
- Modus, Grep, GrepdotAI: Relevant to his emphasis on job-specific benchmarks and deterministic, repeatable code generation from agent workflows.
- Anthropic, OpenAI, frontier models: Broader ecosystem context for his claims that stronger base models reduce old prompt scaffolding while increasing the importance of orchestration layers.
Frequently asked questions
What is Garry Tan?
Investor and startup leader mentioned sharing eval tooling and operational advice for agents and CI workflows. The newsletter links him to repositories and experimental modes used for agentic engineering.
What should AI product managers know about Garry Tan?
Garry Tan is repeatedly cited for practical guidance on agent evals, coding workflows, and production-oriented AI harness design. He shared the gbrain-evals repository with 40+ categories and stressed measuring retrieval, latency, and cost per compute. He described using Capy.ai, GStack, and GitHub/CI workflows to accelerate bug fixing and pull-request execution.
Explore this topic
Newsletter Mentions (57)
“#2 𝕏 Garry Tan shared his gbrain-evals repository, which includes 40+ categories that can each be auto-researched and auto-improved. He said clear evaluations for retrieval, latency, and cost per compute are useful for retrieval and latency data tasks like those performed by GBrain.”
#2 𝕏 Garry Tan shared his gbrain-evals repository, which includes 40+ categories that can each be auto-researched and auto-improved. He said clear evaluations for retrieval, latency, and cost per compute are useful for retrieval and latency data tasks like those performed by GBrain. #6 𝕏 Garry Tan commented that Capy.ai works best with experimental modes, including Captain, turned on.
“Garry Tan commented that lab-developed harnesses are incentivized to burn tokens, giving startup-developed harnesses utility.”
#19 𝕏 Garry Tan commented that lab-developed harnesses are incentivized to burn tokens, giving startup-developed harnesses utility. He said Grep can convert token burn into deterministic, tested, repeatable code by observing agent use, while quoting a GrepdotAI agent workforce startup announcement.
“Garry Tan said he had just deleted about a thousand lines of markdown from gstack because frontier models are now good enough that the old tricks generally aren’t needed anymore.”
GenAI PM Daily - Perplexity Computer now generates interactive 3D learning visualizations October 04, 2026 GenAI PM Daily 🎧 Listen to this brief 3 min listen Today's top 12 insights for PM Builders from X and LinkedIn. Perplexity Computer now generates interactive 3D learning visualizations #6 𝕏 Garry Tan said he had just deleted about a thousand lines of markdown from gstack because frontier models are now good enough that the old tricks generally aren’t needed anymore. #7 𝕏 Garry Tan commented that harnesses can use the same frontier models with better coordination to produce better code outcomes.
“Garry Tan highlighted Capy’s unexpected, model-agnostic cross-session coordination, saying his GBrain collaborator Sina started a fix wave that steered around Tan’s existing work across multiple Capy threads.”
#19 𝕏 Garry Tan highlighted Capy’s unexpected, model-agnostic cross-session coordination, saying his GBrain collaborator Sina started a fix wave that steered around Tan’s existing work across multiple Capy threads.
“Garry Tan highlighted Modus’s open-source test designed to assess whether AI can perform actual junior-auditor work, rather than merely answer accounting questions.”
#11 𝕏 Garry Tan highlighted Modus’s open-source test designed to assess whether AI can perform actual junior-auditor work, rather than merely answer accounting questions. He predicted that every job will eventually have similar evaluation benchmarks.
“Garry Tan recommended the AsideAI browser with MCP for agents facing antibot issues, suggesting it run on a spare laptop or computer kept plugged in.”
#6 𝕏 Garry Tan recommended the AsideAI browser with MCP for agents facing antibot issues, suggesting it run on a spare laptop or computer kept plugged in. He said it lets agents use the user’s real credentials through a real Chromium browser, calling it a “game changer.”
“Garry Tan shared his “favorite way to fix bugs”: using @capydotai with GStack /autoplan on a production issue and GPT-6 medium reasoning.”
#10 𝕏 Garry Tan shared his “favorite way to fix bugs”: using @capydotai with GStack /autoplan on a production issue and GPT-6 medium reasoning.
“Garry Tan said he uses @capydotai daily on all his GStack/GBrain pull requests and is going at least 4x faster than with raw Codex or Claude Code via Conductor.”
#18 𝕏 Garry Tan said he uses @capydotai daily on all his GStack/GBrain pull requests and is going at least 4x faster than with raw Codex or Claude Code via Conductor. He added that it handled something on a recent pull request that out-of-the-box Codex would not have done, though the specific action was not provided.
“Garry Tan shared an example of Capy.ai executing an ambitious bug-fix wave on GBrain, demonstrating clear task delineation, automatic parallelization, and a clean GitHub pull-request and CI workflow.”
#11 𝕏 Garry Tan shared an example of Capy.ai executing an ambitious bug-fix wave on GBrain, demonstrating clear task delineation, automatic parallelization, and a clean GitHub pull-request and CI workflow.
“Garry Tan said every AI harness should support Tailscale, noting that Muse and Grok Bot support it out of the box while Codex and Claude Code cloud containers currently do not.”
GenAI PM Daily September 18, 2026 GenAI PM Daily 🎧 Listen to this brief 3 min listen Today's top 20 insights for PM Builders, ranked by relevance from Blogs, X, and YouTube. Garry Tan said every AI harness should support Tailscale, noting that Muse and Grok Bot support it out of the box while Codex and Claude Code cloud containers currently do not.
Related
An AI company mentioned for releasing a report on model behaviors and for internal tool use in later context. Relevant for PMs tracking safety reporting and frontier model operations.
Anthropic’s coding agent experience referenced as one of the targets for reusable skills. It is part of the multi-agent workflow ecosystem described in the newsletter.
An AI research and product company behind models, tools, and repositories. In this newsletter it is referenced through the openai/math repository.
Anthropic’s conversational AI model family, referenced here as handling a task better than the speaker since around Opus 4.8. It is being discussed in terms of agent capability and workflow delegation.
An AI coding agent referenced as a target platform for reusable software-engineering skills. It is listed alongside other coding assistants in the workflow ecosystem.
Co-founder of LangChain, cited here discussing model routing inside an agent harness. He frames many agent steps as cheaper, typed decisions rather than frontier-model generation.
Google's advanced AI research organization. In this newsletter it is mentioned via SynthID Detector and its Gemini live models.
An AI company associated with Devin, mentioned for allowing users to connect personal ChatGPT plans to the agent. Relevant to AI PMs thinking about consumption-based product design.
A product writer and commentator sharing emerging PM playbooks for zero-to-one work. Here he highlights a hybrid build-and-coordinate approach for AI product teams.
OpenAI’s chat product, referenced here as the user-owned subscription plan connected to Devin. For PMs, this highlights quota-linked integration patterns.
A partner company listed in the Sign in with ChatGPT launch set. The newsletter provides no additional context beyond the integration mention.
A protocol used for connecting tools and agents through standardized interfaces. In this newsletter it appears in the context of servers and tools for managing agent context and integrations.
Google’s AI assistant/model family used here as an enterprise work agent that can handle knowledge, media, coding, and sub-agent orchestration tasks.
An AI search and answer company building agentic workflows. In this newsletter, it is associated with Git-repository-based project organization.
An AI infrastructure company focused on agent tooling and harnesses. The newsletter cites its explanation of where typed yes/no calls fit inside a harness.
CEO and co-founder of Google DeepMind, mentioned here announcing open-source SynthID Bio tools.
A product or system referenced as a retrieval and latency data task target. The newsletter uses it as an example of where strong evaluations around retrieval and cost matter.
A product or integration layer that now includes Perplexity Search. The newsletter only mentions it as the destination where Perplexity Search is embedded.
A Claude model variant being updated with stronger biology safeguards to reduce false positives while still routing dual-use biology requests to higher-safety fallback behavior. Relevant for PMs considering safety tradeoffs and product-surface-specific policy tuning.
The code-hosting and collaboration platform used for pull requests and repository workflows. Here it is part of Warp’s automated engineering loop.
A benchmark or model used as a comparison point for Devin's GPT-6 Astra performance. It is mentioned only as a reference for code quality/cost comparison.
An agent or bot referenced as a benchmark comparison for OpenAI's Dots. It is mentioned as being roughly on par, suggesting a comparable personal agent experience.
A model used in the newsletter as a reasoning and execution engine for product experimentation. It is described as generating daily A/B test ideas and implementing winners for a mobile game economy.
An AI agent environment or product that can host models and persona features. In this newsletter it appears both as a place where Qwen3.8-Max is available and as a tool with a /personality feature.
AGI refers to broadly capable artificial general intelligence. Here it is discussed as becoming usable in 2026 and requiring contextual systems around it to be effective.
A communications platform used here as a runtime/connection endpoint for personal AI demos. It is mentioned alongside WebRTC in a quick setup workflow.
Stay updated on Garry Tan
Get curated AI PM insights delivered daily — covering this and 1,000+ other sources.
Subscribe Free