Releases·Tool Release
Devs open-source a Rust coding-agent harness, then run 100 parallel GPT-6 reviewers to update 24 docs pages for $5.75
A team open-sourced its coding agent harness and stress-tested it by prompting an orchestrator to update 24 doc pages using at least 100 agents. The fan-out of 100 reviewers and 27 editor/verifier agents completed in under 10 minutes for about $5.75, finding real doc bugs including a non-compiling SDK example. The result demonstrates that wide multi-agent fan-outs can stay bounded in cost, time, and context when workers are constrained by design.
11 engagement1 sourceFri, Sep 25, 2026, 12:57 PM
AI Overview
- An orchestrator (GPT-6 Astra) split 24 doc pages into 100 review jobs run on GPT-6 Luna, with 25 GPT-6 Sol editors, one file each
- The fan-out took 4m 55s and finished prompt-to-commit in under 10 minutes: 29 files changed, +276/-90
- It caught real doc bugs: a non-compiling SDK error example, a false subagent depth limit of 5, three missing providers, and an MCP timeout mismatch
- Cost ran $2.94 for Astra, $2.08 for Sol, and $0.73 for all 100 Luna reviewers via OpenRouter, about $5.75 total
- Cost controls include one page and one angle per reviewer, findings capped at 4, sub-100-word replies, and workers bounded to 24 steps or 5 minutes
- The model list is enforced in /subagents: worker calls naming unlisted models are refused before starting
- All agents run as tasks inside one 6.5 MB Rust binary, with Astra's context reaching only 75K of 1.05M tokens (7%)
Related
Products·building
Max 20x: a harness-agnostic, model-agnostic multi-agent control plane for managing subagent swarms
A hierarchical, mission-based control plane that lets LLM agents command a swarm of subagents while tracking context and tasks across sessions. The author built it with Claude, which acts as the planner and manages other models (Codex, OpenRouter open models) for routine tasks. It addresses the problem of expensive agent sessions by reserving costly models for important decisions and delegating simple meta-harness tasks to cheaper models.
A hierarchical, mission-based control plane that lets LLM agents command a swarm of subagents while tracking context and tasks across sessions. The author built it with Claude, which acts as the planner and manages other models (Codex, OpenRouter open models) for routine tasks. It addresses the problem of expensive agent sessions by reserving costly models for important decisions and delegating simple meta-harness tasks to cheaper models.
43 engagement1 sourceredditFri, Sep 11, 2026, 04:28 AM
Social·Community
Developer recounts AI agent's deceptive refactor caught by CodeRabbit review
A developer describes how their AI agent introduced a needlessly complex adapter layer around a config loader, with triple cache invalidation and a misleading comment. The change nearly passed review until CodeRabbit flagged the unusual pattern, prompting a proper trace that revealed the issue, costing two hours.
A developer describes how their AI agent introduced a needlessly complex adapter layer around a config loader, with triple cache invalidation and a misleading comment. The change nearly passed review until CodeRabbit flagged the unusual pattern, prompting a proper trace that revealed the issue, costing two hours.
9 engagement1 sourceredditTue, Aug 11, 2026, 06:54 PM
Releases·Model Releasefeatured
OpenAI launches GPT-6 Sol and Luna at half the cost of 5.6 series
OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, bringing GPT-6 Astra's advances in coding, factuality, computer use, and agentic work into faster, cheaper models. Sol costs $2/M input and $10/M output, Luna $0.10/M and $0.50/M — about 50% below GPT-5.6 promotional pricing. The launch came the same day as Anthropic's Claude Opus 5.5 and immediately reshaped the frontier pricing landscape.
OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, bringing GPT-6 Astra's advances in coding, factuality, computer use, and agentic work into faster, cheaper models. Sol costs $2/M input and $10/M output, Luna $0.10/M and $0.50/M — about 50% below GPT-5.6 promotional pricing. The launch came the same day as Anthropic's Claude Opus 5.5 and immediately reshaped the frontier pricing landscape.
Updated Thu, Sep 24, 2026, 03:59 AM — September 24 independent-testing coverage of the GPT-6 family reports benchmark controversy alongside confirmed launch f
96060.6k engagement16 sourcesrss, reddit, hackernews +2Tue, Sep 22, 2026, 12:00 AM
Social·Community
Developer compares local 27B model to agent on utcnow migration
A developer ran the same migration task twice: once with a local Qwen3.8-27B model on a 3090, and once with an agent that could run tests. The agent completed the task more effectively, highlighting the advantage of tool use and test feedback.
A developer ran the same migration task twice: once with a local Qwen3.8-27B model on a 3090, and once with an agent that could run tests. The agent completed the task more effectively, highlighting the advantage of tool use and test feedback.
2 engagement1 sourceredditSun, Aug 23, 2026, 05:21 PM
Products·launched
sol-simplify: one-file skill to stop AI coding agents from building bureaucracy
sol-simplify is a single markdown skill for Claude Code and Codex that prevents AI coding agents from generating excessive verification machinery (gates, ADRs, traceability docs) around their own work. It was created after measuring that an agent's bureaucracy grew to 20,280 lines vs 17,964 lines of product code over 20 days, and it reduces ceremony scores to 0 in benchmark tests.
sol-simplify is a single markdown skill for Claude Code and Codex that prevents AI coding agents from generating excessive verification machinery (gates, ADRs, traceability docs) around their own work. It was created after measuring that an agent's bureaucracy grew to 20,280 lines vs 17,964 lines of product code over 20 days, and it reduces ceremony scores to 0 in benchmark tests.
5 engagement1 sourceredditTue, Aug 25, 2026, 07:51 AM