Frontier Labs Pull The Agent Loop Behind Managed APIs

OpenAI and Anthropic now ship managed agent harnesses; AWS and Microsoft sell the same loop model-free.

@JoshARosen walks through the shift: OpenAI's new Agents API gives developers the managed Codex harness, and Anthropic makes the same argument with Claude Managed Agents. The labs build model and agent loop together and claim each improves the other: better harnesses because they control the models, better models because they control the harnesses.

AWS and Microsoft sell managed harnesses that do not require their own model underneath: AWS AgentCore Harness and Microsoft Foundry Agent Service. Rosen counts the shared patterns: agents declared as resources, versioned releases with A/B testing, model swaps inside a live session, tools parked in shared infrastructure such as Foundry Toolboxes and AgentCore Gateway, and skills loaded on demand.

Microsoft's Agent Optimizer adds an outer loop that generates and tests new agent configurations against production traffic before a human promotes one. Rosen's closing questions for builders: how much of the agent loop should you still own, and does the harness need to come from the company building the model?

From @JoshARosen

Agents & Coding

Agent Memory Is A Sticky Note, Not A Second Brain

@MichaelGannotti splits Hermes and OpenClaw memory into tiny hot files and an on-demand vault.

@MichaelGannotti argues MEMORY.md is working memory, not a library: Hermes keeps it near 2,200 characters plus 1,375 for USER.md, frozen at session start so the prompt cache stays stable. A vault such as an Obsidian wiki should be searched on demand, never stuffed into prompts. His setup ships with Hermes v0.14: a scoped AgentVault, daily notes for the timeline, wiki pages for knowledge, and a weekly human review.

From @MichaelGannotti

Models & Hardware

DeepSeek V4.1 Flash Runs On Three DGX Sparks At 69 tok/s

@MiaAI_lab posts 69 tok/s single stream on a real agentic task; four Sparks run it better.

@MiaAI_lab reports DeepSeek v4.1 Flash serving 69 tok/s on a single stream for a real agentic task across three DGX Sparks, with the same repo covering three and four box setups. Asked to compare a 2x2 GLM 5.3 Flash cluster against 4x DeepSeek, Mia answers: "DeepSeek v4.1 Flash performance so well on 4 units I'd just run that." The setup runs ring topology with no switch required.

From @MiaAI_lab

Qwen3.8 Flash-Next Recipe Hits 66.17 tok/s On One Spark

@sojufx tuned a production profile: 66.17 tok/s single stream, 179.31 aggregate at 4 streams.

@sojufx publishes the fastest Qwen3.8 Flash-Next profile he found for one DGX Spark after multiple rebuilds: 66.17 tok/s single stream and 179.31 tok/s aggregate at 4 concurrent streams, tuned for a 262K-context production setup with FP8 KV, BF16 recurrent state, MTP K=3 and a 47K draft vocabulary. Built on @MiaAI_Lab's runtime work. He notes the numbers are throughput only, with no formal quality suite.

From @sojufx

Industry

Regulation Odds Jump To 27% On Polymarket

@IterIntellectus calls the surge a capture attempt; @kevinnbass blames a one-sided publicity war.

Polymarket put a 27% chance on U.S. AI regulation after researchers warned advanced AI could "kill us all by the end of the decade," and @IterIntellectus calls the panic "a massive regulatory capture attempt by the largest AI labs." @kevinnbass argues the American public turns pessimist while Chinese state media sells optimism, and urges accelerationists to build "overreach simulations" to counter regulatory capture.

From posts by @IterIntellectus, @kevinnbass

Game Studio Replaces Unreal With Its Own AI-Built Engine

@Grummz says six months of AI coding produced Kun-Zhi and 93x faster development.

@Grummz reports his studio spent 6 months using AI to replace Unreal Engine under the game Em-8ER, and now develops 93x faster on the in-house Kun-Zhi engine. Today's update adds character animation, collision meshes, outfits and pilotable vehicles. The engine ships to backers first; partners include Razer and AMD. In the replies he says he rotates Codex, Grok and Claude, mostly on Astra, and writes in C with some Elixir.

From @Grummz

Voice Clone Scam Call Impersonates A Caller's Own Wife

@Abomination81 describes a cloned-voice scam call and a family verbal password.

@Abomination81 says he took a call in his wife's voice asking for a credit card for gas while she sat beside him, with odd cadence he guesses was a voice filter over a clone. His advice after the scare: agree on verbal passwords with family. Commenters report the same defenses; he thinks scammers may have recorded her voice on a prior call, since it "only takes a few seconds."

From @Abomination81

Sources

Thursday, September 10, 2026 Back to the Archive Saturday, September 12, 2026