$ today_in_10_seconds
Microsoft Research released Orchard, an open framework whose actual contribution isn't a model — it's the environment. Small open-weight agents trained inside it hit 69.7% on SWE-bench Verified with roughly 3B active parameters, and one assistant agent went from 18.6% to 51.5% success purely by being trained inside the harness it would actually run in. Everything shipping this week points the same direction: the scaffolding around the model is where the performance now lives.
$ lead_story
Microsoft's Orchard makes the agent environment a reusable service
Orchard Env is a lightweight Kubernetes-native environment service that creates, manages and tears down thousands of isolated sandboxes in parallel, and it deliberately sits outside any training framework — the same service backs software-engineering, web-browsing and personal-assistant agents without modification, across data distillation, RL rollouts and evaluation. The sharper idea is harness-native training. Capable agents rarely run as a bare model; they run inside stateful multi-process harnesses like Claude Code, Codex and OpenClaw, and open training tooling generally can't handle those — so researchers train on a simplified stand-in and deploy into the real thing, creating a mismatch. Orchard closes that gap with a proxy that records the harness's own model calls as training data while each rollout runs in its own container, letting you train end-to-end inside ReACT, ZeroClaw, OpenClaw or Codex. The results argue for the approach. Orchard-SWE reaches 69.7% on SWE-bench Verified — 73% with value-model reranking — at roughly 3B active parameters. Orchard-GUI is a 4B vision-language browser agent trained on 400 demonstrations plus 2,200 open-ended tasks, averaging 68.4% across WebVoyager, Online-Mind2Web and DeepShop. And under the Codex harness, Orchard-Claw's success rate climbed from 18.6% untrained to 51.5% after training in-harness. Be clear-eyed: 69.7% is state of the art for open models of that size, not frontier — closed frontier coding models currently sit in the high 80s to mid 90s on the same benchmark. The interesting number is the cost curve, not the ceiling.
// why_it_matters: If a 30x smaller model gets to 73% because the environment and harness are right, your agent's bottleneck probably isn't the model you picked.
$ also_worth_attention
AWS quietly retired the managed-agent abstraction
Amazon Bedrock Agents — the November 2023 service — is now Bedrock Agents Classic and closed to new customers as of 30 July 2026, with AWS pointing everyone at Bedrock AgentCore instead. Classic sits in maintenance mode and gets no new features, which in practice means the models available to it today are the models it has forever. The philosophical shift is the story: Classic was a console-configured agent where AWS owned the orchestration loop and your code lived at the edges in Lambdas. AgentCore inverts it — your agent is your code in whatever framework you like, and AWS sells you the runtime, memory, gateway and identity around it. Same conclusion as Orchard, arrived at commercially: own the loop, rent the environment.
link →The harness is now a bigger variable than the model
This is the uncomfortable finding that makes Orchard's design choice look obvious in hindsight: on Terminal-Bench 2.0 the same model can swing 30 to 50 percentage points depending on whether it's wrapped in Claude Code, OpenHands or a homegrown loop. Independent evaluators have responded by standardising on a minimal bash-only harness precisely to push the burden back onto the model rather than the scaffold. Practical upshot: when a vendor or a blog tells you model X is best for agents, the claim is meaningless without the harness, the tool set and the retry policy attached.
link →A land rush is forming around cross-agent memory
Memmy shipped as a local-first memory hub that gives Cursor, Claude Code, Codex, OpenClaw and others one shared persistent context, and it will scan your existing agent history on install to bootstrap that memory. Memorix does the same thing over MCP with the memory living under the Git project rather than inside a chat window, and Zilliz shipped memsearch, a unified memory layer backed by Markdown and Milvus. Note what these all assume: that you use several agents and none of them should be the system of record. That's the same bet Orchard makes about training trajectories — treat them as persistent assets so each generation inherits from the last instead of starting from scratch.
link →
$ bigger_picture
The trend behind the headlines
Four unrelated releases, one direction of travel: value is migrating out of the weights and into the scaffolding. Orchard's own conclusion is that the environment layer matters — make it open and reusable and small models get startlingly far. AWS reached the same place from the other end, killing the managed agent and selling infrastructure instead. The benchmark community is standardising harnesses precisely because harness variance now swamps model variance. And a dozen indie projects are racing to own the memory layer nobody's model provides. If you're building, the strategic question stopped being which model some time ago; it's which parts of the environment, harness, memory and eval loop you own versus rent — because that's now what determines whether your agent actually finishes the task.
mini-swe-agent
Roughly 100 lines of Python for the agent class, bash as its only tool, no tool-calling API required, and it scores over 74% on SWE-bench Verified — built by the Princeton and Stanford team behind SWE-bench and SWE-agent, and used by Meta, NVIDIA and IBM. It's also the framework Orchard-SWE is built on, which makes it the cheapest way to test today's thesis yourself: strip the scaffold to nothing, see what your model actually does.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime