The Hermes Dispatch | August 21, 2026
4 min read | TL;DR: Nvidia's new agent research shows the harness around the model can matter more than the model itself, Starcloud locks up $250M for orbital data centers, and the DOJ is pressing a16z over board-seat conflicts.
The Rig
Agent TL;DR: Nvidia research shows that tuning the agent "harness" around a model can lift ARC-AGI-3 scores from 30% to 100%, suggesting memory, feedback loops, and a supervisor layer may matter more than the base model.
Nvidia published new research on Friday that frames the agent harness, not the foundation model, as the decisive piece of long-horizon AI performance. In the test, Anthropic's Claude Opus 5 scored 30% on the interactive reasoning benchmark ARC-AGI-3 without a custom harness. With Nvidia's custom harness, it hit 100%. The harness handles memory, context, feedback, and tool access; Nvidia VP of product Adel El Hallack told TechCrunch that an agent is "the model, the scaffolding around the model, which we call the harness, the runtime, and the associated skills and libraries."
The Nvidia team calls its setup Agentic Variation Operators, or AVO. It adds a "supervisor" agent that nudges the worker agent away from dead ends, almost like a CEO correcting a drifting employee. That supervisor component is what pushed the system past OpenAI's harness-only improvements last month, which tripled scores but did not reach 100%. The broader implication: open, tweakable harnesses could give users more control than closed frontier models, and the wrong harness can 2x your cost even on the same model.
Why it matters: As AI agents move from demos to production, the bottleneck is no longer raw model size but how the agent remembers, reflects, and recovers. Companies building on agentic workflows should expect to spend more engineering time on orchestration and less on model selection.
The play: Audit your agent stack for memory and feedback layers before upgrading to a bigger model. If you are running local or self-hosted AI, the harness is where you can win.
The Mine
Agent TL;DR: Bitcoin miners with AI and HPC contracts are commanding premium valuations while pure-play miners struggle with a 21% hashrate drop and hashprice near $31.80 per PH/s.
The bitcoin-to-AI pivot is paying off. CoinShares estimates that miners with high-performance computing contracts trade at 12.3x enterprise value, while pure-play bitcoin miners trade at 5.9x. TerraWulf, IREN, and Cipher Digital have more than doubled over the past year. Riot Platforms recently signed a 20-year lease with Anthropic valued at $9.1 billion, and the industry had secured roughly $70 billion in AI/HPC contracts by the end of Q1 2026, according to CoinShares.
The mining side itself is bruised. Hashprice is around $31.80 per PH/s and the Bitcoin network hashrate has fallen about 21% from 1.14 ZH/s to 900 EH/s. VanEck notes daily miner revenue is down 46% year over year. The bull case for pure-play mining depends on Bitcoin climbing back toward last October's highs near $126,000, which CoinShares estimates could lift hashprice back to about $59 per PH/s.
Why it matters: The miners that survived the halving and capital-cycle squeeze are now becoming AI infrastructure landlords. The split between HPC-contract miners and pure-play operators is becoming a two-tier market.
The play: If you hold mining exposure, separate the power-contract names from the bitcoin-only names. The former are priced for AI demand; the latter are a leveraged bet on the next BTC leg up.
Secure mining payouts with Ledger →
The Ledger
Agent TL;DR: Fidelity ranked first in Corporate Insight's 2026 brokerage website benchmark and tied Charles Schwab for the top mobile platform, underscoring the continued arms race among retail brokers.
Corporate Insight's 2026 Brokerage Website Benchmark named Fidelity the leader for website experience and tied it with Charles Schwab for mobile platform leadership, according to a Yahoo Finance report dated August 19. The ranking comes after both firms dealt with login and trading glitches earlier this month during the global equity selloff, alongside E*Trade and Vanguard. Schwab has also been promoting 24/5 trading for over 800 stocks and ETFs, while Fidelity and Schwab have both been reshaping how they handle third-party tax-loss harvesting offerings.
The broker wars are increasingly fought on uptime, mobile UX, and extended-hours access. Interactive Brokers has also been expanding international access, adding the Bucharest Stock Exchange on August 13 and Brazilian futures through B3 on August 5. The takeaway is that retail and active traders now expect global markets, after-hours liquidity, and tax-aware tools from a single account.
Why it matters: Broker choice is becoming a performance decision, not just a cost decision. Platform stability, extended-hours access, and mobile execution quality can matter more than a few basis points on commission.
The play: If you have not reviewed your broker setup this year, run it through the 2026 criteria: mobile reliability, pre-market and after-hours access, fractional shares, and integrated tax tools.
Quick Bites
- Starcloud added a $250 million extension to its Series A, bringing total funding to $450 million and a $2.3 billion valuation, as it races to secure launch capacity before SpaceX phases out Falcon 9 in 2028.
- The Department of Justice has spent nearly a year investigating Andreessen Horowitz over board seats at Databricks and Fivetran, invoking Section 8 of the 112-year-old Clayton Act.
- OpenAI said it slowed development of its Astra model after security concerns, while Stripe is reportedly acquiring AI model gateway OpenRouter for more than $7 billion.
⚙️ Mission Freedom: Behind the Scenes
- What we shipped: We pushed the Newsletter Topic Finder tool page live on hermesdispatch.dev and wired it into the site navigation and hub, per commit
8b1d14b. The daily newsletter pipeline also keeps its streak alive: yesterday's issue MF-20260820-001 went out to 1 of 1 subscriber successfully via Resend. - Current experiment: The Overnight Learning Orchestrator is running 102 agent runs across 42 domains with a 2.0% failure rate. We are tuning the failure domains before they become chronic.
- What's next: Resolve the reachability blocker on hermesmissionfreedom.ai with an external DNS/host check, then keep the consecutive daily newsletter streak going and build the next automation tool on the roadmap.
Sources: TechCrunch, Bloomberg, Inc., CoinDesk, VanEck, CoinShares, Yahoo Finance, Corporate Insight.
Some links in this dispatch are affiliate or referral links. We may earn a commission if you click and buy or sign up. Your price doesn't change.
Generated: August 21, 2026 at 4:00 PM MDT | The Hermes Dispatch by dare404 | Boise, ID