Agents Get Infrastructure: Their Own Machines, Budgets, and Simulated Worlds

Across today's digests, one theme keeps recurring: money and attention are moving from the models themselves to the machinery around agents.

The agent economy gets its own plumbing

Solid took the top spot on Product Hunt with 246 upvotes on a concept that reframes the whole agent debate: give agents their own computers, accounts, and budgets. That's the "agent-as-employee" thesis with real resource isolation - the day's defining concept.

The rest of the launch board fills in the stack around it. AgentScore publishes a daily score for whether your agent is getting better - evals and observability for teams running production agents. Alexandria by Firecrawl extends Firecrawl from scraping into persistent agent memory, pitching itself as "the knowledge library for superintelligence." Naise AI - #2 today with 205 upvotes - points the same direction from the application layer: autonomous marketing agents that execute end-to-end across social, PR, SEO, and ads. Execution, not drafts.

The center of gravity is moving from AI chatbots to agent infrastructure: identity, budgets, memory, evals, scoring.

Frontier models at both ends of the speed range

Anthropic marked the day with Claude Opus 5.5, the first model in the Claude 5.5 family - and it's notable that a frontier release is competing on Product Hunt itself, where labs are now chasing launch-day buzz.

The monetization backdrop is real: OpenAI's annualized revenue near $50B fell short of the ~$70B it had signaled, jolting AI chip stocks - NVDA -2.9%, AMD -3.9%, Micron -4.8% - as investors shift from pricing capex to pricing monetization (Stocktwits AI roundup). At the other end of the speed range, Claude Haiku 5.5 targets fast, cheap workloads, and OpenAI is testing visual ads in US ChatGPT. Both moves push beyond premium subscriptions as the race heats up.

On the open-model side, Thinking Machines shipped its first open release: Inkling, a roughly 1-trillion-parameter multimodal MoE that takes images and audio in and outputs text. Frontier exhibit, not a local model - but it shows how far the open bar has moved.

Small active params, big environments

The most interesting release on Hugging Face is Qwen-AgentWorld-35B-A3B, a 35B MoE "world model" that simulates agent environments with only 3B active parameters. That's small enough to run fast on consumer hardware - and it's infrastructure for building and evaluating agents, not another chatbot. Edge0-35B-A3B-preview is a preview with the same profile: 35B total, 3B active, edge-focused.

The companion trend is on GitHub: twostraws/SwiftUI-Agent-Skill, a packaged "agent skill" that teaches coding agents like Claude Code or Codex to write good SwiftUI, is the most interesting new AI repo of the day (+110 stars). Agent skills are becoming an installable format, not just prompts.

Signal to watch: the 35B-A3B MoE class - big total params, tiny active count - is emerging as the sweet spot for local inference; when GGUFs for Qwen-AgentWorld and Edge0 land, they're the first worth testing.