Show HN: Kodiqa Agent
An agent client pitched as "one agent, every model, zero limits" — a single interface across providers.
7 articles · 3 categories
What happened in AI — Jun 6, 2026
Saturday, Jun 6, 2026
In 30 seconds
- Latent Space called it a quiet day of recursive self-improvement — not much happened on the frontier.
- Weekend Show HN leaned into agent control: Kodiqa (one agent, every model), Hermes (community web UI), and Sub-Agent MCP for delegation.
- Agent-ML-skills offers a skills pack to stop Codex/Claude/Cursor from repeating common ML mistakes.
- An engineering essay reframes every AI agent feature as a cache-invalidation surface.
- Simon Willison sandboxes Python by running MicroPython compiled to WASM.
A quiet Saturday on the frontier — Latent Space's daily dispatch again shrugged with "not much happened today" — but a busy weekend for builders shipping agent tooling. The day's Show HN feed was all about controlling and orchestrating agents: a one-agent-every-model client (Kodiqa), a community web UI (Hermes), MCP-based sub-agent delegation, and a skills pack to teach Codex, Claude, and Cursor to stop making the same ML mistakes.
The sharper reading came from two engineering notes: an argument that every AI agent feature is really a cache-invalidation surface, and Simon Willison's latest take on sandboxing — running Python safely via MicroPython compiled to WASM.
Agent Tooling & Orchestration 4 items
The weekend's loudest thread: clients, UIs, and orchestration layers for running and delegating across agents.
Hermes – Community Web UI with Agent
A community-built web UI with an integrated agent for interacting with models in the browser.
Show HN: Sub-Agent MCP: LLM delegation and sub-agent orchestration via MCP
An MCP server for LLM delegation, letting a primary agent spawn and orchestrate sub-agents through the protocol.
Agent-ML-skills – Teach Codex/Claude/Cursor to stop making ML mistakes
A skills pack that gives coding agents like Codex, Claude, and Cursor guardrails to avoid common machine-learning mistakes.
Engineering Notes 2 items
Two builder-focused reads on the hard parts: state invalidation in agents and safely sandboxing untrusted code.
Every AI Agent Feature Is a Cache Invalidation Surface
An essay arguing that each new agent feature really introduces another cache-invalidation problem — the perennial hard problem resurfacing in agent design.
Running Python code in a sandbox with MicroPython and WASM
Simon Willison's latest sandboxing approach — running Python via MicroPython compiled to WASM — which he says finally hits the properties he's been after.
Commentary 1 item
The day's signal-to-noise read.
[AINews] not much happened today
Latent Space's daily dispatch calls it a quiet day of recursive self-improvement — a useful signal that the week's big moves landed elsewhere.