Ask HN: What's your current AI agentic setup?
AI power users reveal their bleeding-edge agentic setups, showcasing a diverse landscape of paid subscriptions, custom tooling, and sophisticated workflows. This deep dive into daily AI integration highlights a significant shift in developer productivity, even as users grapple with cost, privacy, and the evolving nature of human-AI collaboration.
The Lowdown
The "Ask HN" post initiated a fascinating self-reported survey among Hacker News users, delving into their current AI agentic setups. The author, driven by a personal observation of how their workflow had changed dramatically, sought a snapshot of common practices, tools, and challenges. The questions aimed to uncover:
- AI Subscriptions: Which commercial AI models (e.g., Claude, ChatGPT/Codex) are users personally paying for, including specific tiers and multiple accounts.
- External Models: Utilization of pay-as-you-go APIs, OpenRouter, or locally hosted models, specifying providers and model types.
- Coding Harnesses & Interfaces: The preferred tools for interacting with AI for coding, ranging from default CLIs and IDE integrations to specialized harnesses like OpenCode, Pi, or custom-built solutions.
- Agent Orchestration/Swarm Tools: Adoption of advanced systems for coordinating multiple AI agents, including named examples and bespoke solutions.
- Broader Agent Platforms: Usage of comprehensive platforms like OpenClaw, Hermes Agent, or Grok bot, and their primary use cases.
The resulting discussion painted a vivid picture of a rapidly evolving field where users are actively experimenting, customizing, and pushing the boundaries of AI integration in their daily work and personal lives, reflecting both immense gains in productivity and emerging complexities.
The Gossip
Subscription Stew & Model Medleys
Users often combine multiple paid AI subscriptions (Claude, ChatGPT/Codex, Gemini) with more cost-effective API-based models (DeepSeek, GLM, Qwen via OpenRouter) or local LLMs. The goal is to balance performance, cost, and specific model strengths for various tasks, leading to frequent discussions about optimizing spending and navigating usage limits.
Orchestration Innovations & Harness Hacks
A significant number of users are building their own custom harnesses, UIs, and orchestration tools, driven by a need for tailored workflows and dissatisfaction with off-the-shelf solutions. Common base tools like 'pi', 'OpenCode', 'Herdr', and 'T3 Code' are frequently extended, enabling seamless integration, advanced task management, and multi-agent coordination.
Productivity Proliferation & Existential Reflections
AI agents are radically transforming workflows, automating everything from coding and project management to personal errands. While many report vastly increased productivity and the ability to focus on higher-level tasks, there's an undercurrent of concern about 'AI addiction,' the changing nature of creative work, and a potential loss of direct engagement with the craft.
On-Premise Principles & Provider Perils
Privacy, vendor lock-in, and the unpredictable nature of cloud providers (e.g., account cancellations) are pushing many towards local LLM inference and open-source models. Users invest in powerful personal hardware or seek 'zero data retention' APIs, valuing control over their data and tools, even if it means overcoming performance or configuration challenges.
Agentic Ambitions & Simplicity's Success
While some pursue complex multi-agent orchestration, others find success in simpler setups, arguing that modern frontier models are capable enough for direct interaction, with subagents spawning autonomously when needed. This highlights a debate over whether advanced agentic systems are truly beneficial or add unnecessary complexity and token consumption, a form of 'bikeshedding' in the new AI paradigm.