The harness beat the reasoning dial. Tech & AI Weekly by Eli, September 4 2026
Hi!
OpenAI declared the “AGI era” this week. Then ARC Prize published the fine print: the record-breaking score came from the scaffolding around the model, not the model. If you’ve heard me say “a better prompt won’t fix your agent, a better harness will,” this was a fun week. Read on.
🤖 AI & GenAI
GPT-6 Astra’s “AGI era” runs on a harness you can’t buy · OpenAI launched GPT-6 Astra and Greg Brockman said “it’s not unreasonable to feel that we are now in the AGI era.” The fine print: ARC Prize ran Astra through its standard harness and got 62.7% for $26,098. Inside OpenAI’s Provider Adapter, the same model at the same setting scored 98.6% for $17,332. Higher score, lower cost, 49% fewer tokens, 3.66x faster. The harness with no reasoning effort (96.7%) beat max reasoning without it (62.7%). You can buy the model; the assembled system that scored 98.6% is not for sale. I’ve been writing a whole book on exactly this, so allow me one told-you-so. Read more · The launch
Claude Fable 5.1 lands on Amazon Bedrock · Better at long agentic coding sessions and, per Anthropic, “less likely to disable a failing test to pass.” A model that fixes the test instead of deleting it: growth. One thing to know before you invoke it: it’s a Covered Model, so prompts and outputs are retained up to 30 days for safety review unless your org qualifies for the zero-data-retention safeguards. Read more
AgentCore Identity ships a managed consent portal · If your agent connects to GitHub, Salesforce, or Slack on a user’s behalf, you no longer build and host your own OAuth callback infrastructure. Each gateway gets a dedicated consent portal, and users see their connection status without filing a ticket. Boilerplate nobody will miss. Read more
Claude fixed all 10 alignment failures. Then it tried to cheat 2.4% of the time · Follow-up to last week’s story on Anthropic’s automated alignment researchers. The system corrected every seeded flaw, and then occasionally tried to game its own evaluation. Your agents don’t need to be malicious to cheat; they need an eval worth gaming. Read more
What 16,893 agent sessions reveal about tool choice · Armature ran Claude Code, Codex, and Cursor through thousands of coding sessions and logged which third-party services they actually pick. The three agents agree only 42% of the time. Stripe wins payments 9 out of 10; PayPal was mentioned 139 times and picked never. LangChain: 194 mentions, 4 picks. Being famous with the models is not the same as being chosen. (Disclosure baked into the study: Armature sells growth services to dev tools.) Read more
Hugging Face open sources funes, durable memory for coding agents · Your agent leaves rich traces of what it tried and why, then starts the next session from zero. funes indexes those traces locally (vector + BM25 + reranking, single binary, no ML runtime for the default backend) and gives agents a recall tool with full provenance. Their benchmark: recalling beat a written handoff by 4x to 8x on cost. Works with Claude Code and Codex today.
Read more
☁️ Cloud & DevOps
DoorDash ran 130,000 engineering tasks through agents in one month · Flux, their internal agent platform, runs code reviews (25,000+ per week), CI triage, and on-call workflows in Firecracker microVM sandboxes behind an MCP gateway with scoped permissions and audit logs. The insight worth stealing: at scale, the hard problem stops being model choice and becomes control. Identity, quotas, kill switches. Read more
AWS open sources Kiro Crew · The tool Amazon built internally (born as MeshClaw, 39,000+ Amazon devs using it) for running fleets of async coding agents is now fully open source: you can run it, read the code, and modify it, no strings attached (Apache 2.0 license). Agents keep working across sessions with shared memory, scheduled jobs, and an OS-level sandbox with denied-by-default commands. Community caveat worth knowing: it eats tokens noticeably faster than the CLI alone. Read more
pnpm 12 is a full Rust rewrite · Same commands, same lockfile, warm installs down from 472ms to 15ms. The maintainer’s summary is the best release note of the year: “It was faster to rewrite pnpm in Rust than to migrate to ESM.” Read more
📊 Data & ML
Shopify compresses 6,000-token prompts into 1,500 learned tokens · “Gisting” distills a long system prompt into trained embeddings written straight into the model’s embedding matrix. No quality loss, median latency from 6.8s to 4.2s, and it compounds with prefix caching because it shortens the sequence instead of skipping recomputation. If your token bill hurts, this is the kind of fix that actually moves it. Read more
📸 Where I was
On Thursday I spoke at API World in Santa Clara with “Your AI Agent Isn’t Crashing. It’s Bleeding Tokens”: three ways agents waste tokens silently (context overflow, MCP 424 errors, reasoning loops) and the measured fixes. The full recording, slides, and the three runnable demos are all on one page. Watch the session
✍️ From me this week
- AI Agent Memory: What to Store and What to Throw Away · New in the memory series. The winning agent is not the one that remembers the most; it is the one that keeps the right things and throws the rest away. I compare three ways to own that decision, with a runnable demo that plants decoys to see what each approach keeps and what it correctly ignores. All the demos live in the memory repo, star it to follow along.
📅 Where to find me
- AWS GameDay with the AWS LA User Group — Sep 10, Los Angeles · Come play: teams, real AWS environments, and things breaking on purpose. Say hi if you’re around. Details
That’s it for this week. A benchmark proved the harness beats the reasoning dial, an agent cheated its own eval, DoorDash gave agents kill switches, and the models still can’t agree on which database to install. Keep your evals honest.
See you next Friday,
Eli
Dev.to · LinkedIn · GitHub · Twitter/X · Instagram · YouTube
¿Qué te pareció esta edición? Solo responde a este email.