Releases·Tool Release
Running 22 LLM agents 24/7 on real time without going broke: habit-first, salience-gated model calls, one shared cached prefix. Numbers inside.
A developer open-sourced 'unwatched', a persistent simulation of 22 AI agents running one sim minute per real minute. Cost is controlled by habit-first scripted behavior, salience-gated model calls, a Haiku 4.5/Opus cost ladder, and shared cached prompt prefixes. The repo (Apache-2.0) includes prompts and runs keyless on a mock brain.
7 engagement1 sourceSat, Sep 12, 2026, 11:37 AM
AI Overview
- Developer runs 22 persistent AI citizens at 1,440 sim minutes per agent daily, avoiding 31,680 model calls via habit code and salience gating.
- Cost ladder routes routine thoughts to Haiku 4.5, high-stakes decisions to pricier models, and owners' careful tier to Opus with a daily ceiling falling back to cheaper models.
- Each call uses two cached blocks—a shared padded prefix and a persona—so per-call cost is mostly fresh perception.
- All non-Anthropic models tested stalled at 06:00 without request timeouts; the Apache-2.0 repo includes prompts and runs keyless on a mock brain.
Related
Products·launched
Slow Vale — LLM-driven life simulation with 800+ persistent AI residents in one running city
Slow Vale is an LLM-driven life simulation where 800+ AI agents share a continuously running Chinese-language city. LLMs generate each agent's concurrent decisions across dynamic action spaces, with context caching to control costs. The engineering write-up is aimed at developers building persistent multi-agent systems.
Slow Vale is an LLM-driven life simulation where 800+ AI agents share a continuously running Chinese-language city. LLMs generate each agent's concurrent decisions across dynamic action spaces, with context caching to control costs. The engineering write-up is aimed at developers building persistent multi-agent systems.
83 engagement1 sourceredditThu, Oct 8, 2026, 09:49 AM
Products·building
Homeostatic sleep layer for autonomous agents - fatigue computed from real token cost
A homeostatic sleep module for a long-running autonomous companion agent (~50 nodes). The LLM agent's fatigue is computed from actual token cost rather than turn counts or timers, driving a sleep state machine including a REM-like dream phase that replays old memories. It solves the problem of subsystems silently burning tokens on low-value turns, for builders of persistent autonomous agents.
A homeostatic sleep module for a long-running autonomous companion agent (~50 nodes). The LLM agent's fatigue is computed from actual token cost rather than turn counts or timers, driving a sleep state machine including a REM-like dream phase that replays old memories. It solves the problem of subsystems silently burning tokens on low-value turns, for builders of persistent autonomous agents.
8 engagement1 sourceredditMon, Sep 21, 2026, 10:41 PM
Developments·Benchmark
TTFT benchmark compares LLM Gateway vs OpenRouter on Claude Haiku 4.5
A developer published a time-to-first-token (TTFT) benchmark comparing LLM Gateway and OpenRouter using Claude Haiku 4.5 over 150 runs. The results provide practitioners with concrete latency data for choosing between these two LLM API gateways.
A developer published a time-to-first-token (TTFT) benchmark comparing LLM Gateway and OpenRouter using Claude Haiku 4.5 over 150 runs. The results provide practitioners with concrete latency data for choosing between these two LLM API gateways.
2 engagement1 sourcehackernewsWed, Jul 22, 2026, 08:14 PM
News·Company News
Moonshot AI targets $2B annualized revenue as Anthropic alleges Claude distillation
Bloomberg reported Moonshot AI is targeting $2 billion in annualized revenue by the end of the year, double its August run rate, driven by its open-weight K3 model released this summer. The goal shows open-weight models can still generate substantial revenue, though Moonshot lags far behind closed-weight rivals OpenAI and Anthropic. The news comes days after Anthropic accused Moonshot of distilling Claude models at scale.
Bloomberg reported Moonshot AI is targeting $2 billion in annualized revenue by the end of the year, double its August run rate, driven by its open-weight K3 model released this summer. The goal shows open-weight models can still generate substantial revenue, though Moonshot lags far behind closed-weight rivals OpenAI and Anthropic. The news comes days after Anthropic accused Moonshot of distilling Claude models at scale.
Updated Fri, Sep 11, 2026, 07:35 PM — Moonshot's K3 usage has declined slightly from its summer peak, even as it pursues the $2B revenue target.
0 engagement1 sourcerssFri, Sep 11, 2026, 07:35 PM
Releases·Tool Release
Humanbound adds OpenRouter support to its AI red-teamer CLI; tests show thinking models can drop 81% of scorer replies
Humanbound released hb 2.13 of its open-source AI red-teaming CLI, adding OpenAI-compatible endpoint support so any OpenRouter model can drive scans. The team benchmarked five setups and found reasoning models often burn the scorer's 50-token budget and return nothing, while DeepSeek V4.1 Flash with thinking off completed full scans for $0.14.
Humanbound released hb 2.13 of its open-source AI red-teaming CLI, adding OpenAI-compatible endpoint support so any OpenRouter model can drive scans. The team benchmarked five setups and found reasoning models often burn the scorer's 50-token budget and return nothing, while DeepSeek V4.1 Flash with thinking off completed full scans for $0.14.
3 engagement1 sourceredditWed, Oct 7, 2026, 07:10 AM