Field guide / sub-domain
11
sources in this field
Updated September 20, 2026
Current thesis
The shortest path to orientation.
This synthesis records claims and practices from the cited sources; reported outcomes and product capabilities have not been independently verified.
Agent frameworks crystallized around specialized primitives—Pipecat (voice), browser-use (web), Mem0 (memory), Composio (1,000+ OAuth), RAGFlow (retrieval), Dify (workflows), Mastra (TypeScript, 1.77M npm downloads)—while security and research package as plugins. MCP is essential; agent-native CLIs with dry-run previews outperform instruction-heavy MCPs. Subrouter, a local Go proxy, routes Codex/Claude Code traffic across subscriptions and API keys via sticky conversation-to-account assignment, selecting accounts by most-constrained usage window and protecting low-headroom accounts. It deploys via LaunchAgent, systemd, or binary with daily autoupdates and defaults security to 127.0.0.1. Major SaaS platforms are now exposing data natively to agents: Stripe released an official MCP server letting agents query live data to build custom reports on MRR, churn, and retention—gated behind Stripe Sigma, targeting advanced analytics users rather than all Stripe customers.
The json-render source adds catalog-constrained UI generation across renderers, with experimental Jev composition kept separate from stable APIs.
Evidence board
Claims worth carrying forward01Thinking Machines made its Inkling model free on OpenRouter, but access is restricted to agentic harnesses only (not general chat use), running for a few weeks starting now.
02Thinking Machines' stated goal for the free Inkling access is to observe real-world agentic performance and use the resulting usage data—disassociated from user accounts—to improve the model.
03Pattern: AI labs are using time-limited, use-case-restricted free API access (e.g., agentic-harness-only) as a deliberate data-collection strategy rather than pure monetization or general-availability launches.
04Stripe released an official MCP server letting AI agents query live Stripe data to build custom reports on metrics like MRR, churn, and retention—an example of a major SaaS platform exposing financial data natively to agent tooling rather than via dashboards.
05Stripe's agent-analytics MCP is gated behind Stripe Sigma (Stripe's SQL-based reporting add-on), meaning the feature targets businesses already paying for advanced analytics rather than all Stripe users.
06Subrouter is a local Go proxy that routes Codex/Claude Code traffic across multiple ChatGPT Pro/Claude Max subscriptions and API keys, using sticky conversation-to-account assignment so cached context stays useful.
07Subrouter's account selection policy scores accounts by most constrained usage window, protects low-headroom accounts for new sessions, spends quota resetting soonest, and prefers subscription OAuth accounts over API-key accounts when tied.
08Security defaults: Subrouter binds to 127.0.0.1 unless explicitly exposed; non-loopback deployments require either an admin token (Bearer/X-Subrouter-Admin-Token header) or Tailscale-based identity auth (--tailscale-auth) verified via tailscale whois against the caller's WireGuard-authenticated peer.
Adjacent fields
Latest evidence
Thinking Machines
Thinking Machines is making its Inkling model freely available on OpenRouter, restricted to agentic harnesses only, for a few weeks to gather real-world agentic behavior data (anonymized) to improve the model.
Stripe
Stripe introduced an MCP (Model Context Protocol) integration allowing AI agents to query Stripe data directly, enabling custom reports on MRR, churn, and retention. The feature requires Stripe Sigma access.
Ассизи
Subrouter is a Go-based local proxy (manaflow-ai) that distributes Codex and Claude Code traffic across multiple ChatGPT Pro/Claude Max subscriptions and API keys, using sticky conversation-to-account assignment, usage-aware selection, multi-tenant isolation, and optional full-transcript recording.
Matt Carey
Cloudflare open-sourced Cloudflare OS, an internal-first platform giving every employee an agent workspace grounded in company context, with a new security/governance framework (Gatekeepers) and a system for building shareable, modifiable full-stack apps as Workers.
Yakko
Railcode is a platform for building internal apps and agents connected to company data, developed locally with Claude Code/Codex and deployed with built-in infrastructure (db, file storage, email, connectors) and granular permissions.
Cathryn
Cathryn Lavery argues the highest-leverage first AI loop isn't a more autonomous agent — it's a system that reads your own Claude Code/Codex session transcripts as evidence, surfacing repeated corrections, tool failures, and workflows worth turning into content, config fixes, skills, hooks, or slash commands. She open-sourced 'agent-improvement-loop,' a local tool that scans sessions, stages proposals into seven categories, and requires manual approval before any change — pairing with a companion piece on why agent-native CLIs beat official APIs/dashboards for a second, non-human user class.
Fivos Aresti
A GTM services company describes a 5-layer architecture (markdown company OS, self-updating context repos, 25+ MCP tool connections, self-improving Pinecone-backed memory, and operating principles) that decouples output from headcount, reportedly reaching 7-figure revenue in 6 months with a fraction of traditional agency headcount.
Andrew Ng
Andrew Ng and Rohit Prasad launched OpenWorker, an open-source, local-first AI agent that produces finished deliverables (documents, Slack replies, calendar updates) rather than just chatting, built on the aisuite library, model-agnostic across major providers and local models, with approval gating, 25+ tool integrations, and scheduled automation.
Nathan Baschez
> @jxnlco I gotchu fam https://t.co/J3YOOpL5ES - Roughdraft.md provides a local-first Markdown review app specifically designed for collaborating with coding agents — relates to [[ai-agents/interaction]] - The tool enables commenting and suggesting edits on Markdown files in a workflow optimized for
Simon Last
> 1/ Some things I've learned recently running coding agents on large-scale projects. Most of this contradicts advice from 6 months ago! - Recent experience with coding agents on large-scale projects reveals that best practices have evolved significantly in the past 6 months — relates to [[ai-agents