新的创业机会，正在 Claude Code 和 Codex 周围长出来｜对谈 Treg

Summary

This 十字路口Crossing interview follows Jason, 石天鑫, and Trigg through three startup directions: Super Design, Coolit, and an agent-facing data and tool marketplace. Their operating conclusion is that stronger general agents compress thin vertical wrappers but create new demand for real-time data, professional APIs, cross-agent tool portability, and interfaces for supervising many sessions.

The episode also qualifies full-automation narratives. The team reports that support, bug repair, monitoring, and other narrow deterministic work can be automated more readily than growth, innovation, channel discovery, or product judgment. Its practical thesis is therefore Human-Agent Collaboration rather than a near-term zero-human company, with Claude Code and Codex treated as distribution ecosystems on which smaller companies can build.

Key Claims

• The team moved from the design agent Super Design to the growth agent Coolit and then to Trigg, using short experiments to decide when model progress or weak results had invalidated a direction.
• Super Design reportedly gained more than 6,000 GitHub stars through open-source distribution, but the team later concluded that general models and coding agents were absorbing more frontend and design capability.
• A roughly two-month attempt to operate Super Design with AI found support and bug fixing substantially more automatable than growth, where channel timing, emerging keywords, competition, and taste still required human judgment.
• A Codex plugin reportedly changed Super Design’s growth curve without additional promotion and brought lower-churn users, illustrating the distribution leverage and dependency risk of an early platform marketplace.
• Trigg aggregates real-time data, professional APIs, and generative-media models behind one key, usage-based billing, and reusable access from multiple agent harnesses.
• The marketplace thesis is that no single supplier is best on coverage, accuracy, and price; task-level feedback from agents could help route later requests toward better tools.
• Small, single-responsibility loops are presented as safer operational automation than broad autonomous management, with monitoring, recovery, and multiple loops used to contain failure.
• Heavy users already run many sessions in parallel, creating demand for task tracking, context transfer, cloud/local coordination, and a primary-agent or “captain and first officer” routing pattern.
• The guests argue that coding speed is becoming less scarce while problem selection, product judgment, cross-domain ability, experimentation, and result review remain differentiating human work.

Key Quotes

“零人公司不是现实场景” — the team’s six-to-twelve-month boundary after trying to automate Super Design operations.

“创新无法依靠自动 Loop 完成” — the distinction between open-ended innovation and narrower operational automation.

“如果没有，就应该考虑转向” — Jason’s test for pivoting when no credible experiment could change the current result.

Connections

• Jason, 石天鑫, and Trigg - founder, founding engineer, and current company at the center of the interview.
• Super Design and Coolit - prior product directions used to test design-agent distribution and automated growth.
• Claude Code, Codex, Anthropic, and OpenAI - general-agent ecosystems and model-company platforms around which the team now builds.
• AI Application Survival Strategy and Model Provider Tool Competition - pivot discipline and the pressure stronger models place on shallow application layers.
• Agent Marketplace and Coding Agent As Universal Action Layer - agent-facing tool aggregation and the coding-agent entry point for non-coding work.
• Human-Agent Collaboration, Routine Agent Automation, and Multi-Agent Collaboration - human judgment, narrow operational loops, and parallel-session coordination.
• Agent Harness, AI Skills, Agent-Facing Interfaces, and Model Workflow Fit - harness, capability packaging, interface, and model-selection layers discussed throughout the episode.

Contradictions

• No settled contradiction was adopted. The source qualifies strong One-Person Company or zero-human-company narratives by arguing that automation concentrates rather than removes judgment, growth, review, and accountability.
• The title says “Treg,” while the body consistently names the company “Trigg.” This ingest uses Trigg for the canonical entity and preserves the title wording rather than treating the spellings as separate companies.
• GitHub stars, revenue multiples, retention, user counts, supplier coverage, feedback quality, model comparisons, security outcomes, and automation results are founder-reported and not independently verified by the supplied summary.
• Trigg had reportedly been public for only about three weeks. Its neutral-marketplace position, usage-based unit economics, supplier quality, cross-agent portability, and agent-feedback data moat therefore remain early hypotheses rather than established durable advantages.