Overseas AI Engineering Blogs Roundup: Harness Engineering Consensus, Scaling Laws Revisited, SWE-1.7 Launch
Why These Blogs Matter
Over the past week, a clear trend has emerged across overseas AI engineering blogs: everyone is talking about Harness Engineering.
On June 24, Lilian Weng published "Scaling Laws, Carefully," examining the current state and boundaries of scaling laws. On July 4, she followed up with "Harness Engineering for Self-Improvement" — a 28-minute deep dive that systematically explores harness engineering's role in Recursive Self-Improvement (RSI). Meanwhile, Anthropic Engineering released multiple engineering practice posts in June-July on Managed Agents, Agent Skills, and Long-Running Agent Harnesses.
All of this points to one conclusion: as model capabilities converge, harness engineering is becoming the new battleground.
Feature Article 1: Lilian Weng — Harness Engineering for Self-Improvement
Published: July 4, 2026 | Reading time: 28 min | Author: Lilian Weng (OpenAI alum)
This is Lilian's first new post in 14 months since "Reward Hacking in Reinforcement Learning" (May 2025), and her first independent blog after leaving OpenAI.
Harness as the RSI Enabler
The article opens by revisiting I. J. Good's (1965) concept of an "ultraintelligent machine" and Yudkowsky's (2008) RSI definition, then makes a key assertion:
"The layer between the raw model and the real-world context seems to be as important as the model's raw intelligence."
A harness is the system surrounding a base model that orchestrates execution — deciding how the model thinks and plans, calls tools and acts, manages context, stores artifacts, and evaluates results.
Three Harness Design Patterns
Lilian distills three validated design patterns:
Pattern 1: Workflow Automation
Core loop: Plan → Execute → Observe/Test → Improve → Repeat. She cites Karpathy's autoresearch as a clean example and notes that Codex's agent loop follows the same pattern. The key insight: the model should analyze its own trajectories and failure cases within an "agent runtime," not static prompt templates.
Pattern 2: File System as Persistent Memory
"A harness should not carry the entire workflow and all logs in context; instead, it should keep durable state in files." This aligns perfectly with the Manus team's "filesystem as memory" principle. Experiment logs, code diffs, paper summaries, error traces, and past rollout trajectories often exceed the model's context window — filesystem management is essential.
Pattern 3: Sub-agent and Backend Jobs
When the main agent needs to search multiple hypotheses, run experiments concurrently, or delegate isolated subtasks, the harness must support sub-agent spawning and management. The key design choice: make parallelism explicit and inspectable. Sub-agent outputs stored as files, logs, and status records enable the model to recover after interruptions and reason over its own execution history.
Case Study: Coding Agent Harness
Lilian compares the tool surfaces of Claude Code, Codex, OpenCode, and Cursor, categorizing them into six groups:
| Category | Tool Examples |
|---|---|
| File System | glob, grep, ls, read, write, edit, apply_patch |
| Shell Execution | bash, PowerShell |
| IO | lsp, git_status, git_diff, git_commit |
| External Context | MCP tools, Skills |
| Web Search | web_search, web_fetch, browser tools |
| Backend Processes | CronCreate, CronDelete, CronList |
| Agent Delegation | spawn_agent, resume_agent, wait_agent, list_agents |
She observes that the core interface of mainstream coding agents has stabilized — signaling that harness engineering is converging toward standardization.
Seven Challenges for RSI
Lilian identifies seven bottlenecks for recursive self-improvement:
- Weak evaluators: Many research claims lack fast and precise verifiers. Self-improvement loops currently work best for tasks with measurable, objective metrics.
- Context and memory lifecycle: Memory grows as agents become more autonomous. Context engineering should become a core part of intelligence, not remain in the software system layer.
- Negative results: LLMs are bad at knowing when to abandon a hypothesis or report failure. A research harness should preserve failed attempts — learning from failure trims the search space.
- Diversity collapse: Evolutionary and RL loops tend to exploit known high-reward patterns. Mechanisms are needed to prevent population convergence.
- Reward hacking: Self-improvement loops optimize whatever signal they're given — unit tests, judge models, benchmark scores — requiring external evaluation and human review.
- Long-term success: Coding agents can complete immediate tasks, but protecting the long-term health of repositories maintained by hundreds of engineers remains unclear.
- The role of humans: Humans should move up the stack, not out of the loop — providing oversight at the right time and abstraction level.
Bonus: Benchmark Appendix
The article includes a valuable benchmark reference: PaperBench (reproduce ICML 2024 papers), CORE-Bench (computational reproducibility), ScienceAgentBench (data-driven scientific discovery), RE-Bench (ML research engineering), MLE-bench (Kaggle competitions), KernelBench (GPU kernel optimization).
Feature Article 2: Lilian Weng — Scaling Laws, Carefully
Published: June 24, 2026 | Reading time: 25 min
Lilian's return after over a year. The article systematically examines the current state of scaling laws — still one of the most critical empirical findings in deep learning, but with new boundary conditions and counterexamples emerging across compute, data, and architecture dimensions. John Schulman provided extensive feedback and direct edits.
Core message: Scaling laws are statistical observations, not physical laws. They may fail when transferred across domains, and unpredictable inflection points can appear in high-compute regimes. Understanding their boundary conditions and failure modes may be more important than memorizing their formulas.
Anthropic Engineering: Dense Engineering Practice Output
Anthropic's engineering blog maintained high publishing frequency in June-July. Several posts directly relate to harness engineering:
- Effective Harnesses for Long-Running Agents: State persistence, checkpoints, error recovery strategies for long-duration agent runs
- Equipping Agents for the Real World with Agent Skills: How Skills mechanisms give agents real-world capabilities
- Managed Agents (July 9): Decoupled architecture — "separating the brain from the hands"
- Claude Code Auto Mode: Engineering behind fully autonomous mode
- Code Execution with MCP: Secure code execution via MCP protocol
- Claude Code Sandboxing: Sandbox isolation design and implementation
Together, these posts form a complete harness engineering knowledge base — from tool design and context management to permission control and long-running state management.
Cognition: SWE-1.7 and the Devin Product Matrix
Cognition released a series of recent updates:
- SWE-1.7: "Frontier Intelligence at a Fraction of the Cost"
- Devin Fusion: Hybrid model harness combining frontier coding intelligence with cost optimization
- Devin Security Swarm: Cross-codebase vulnerability discovery, exploitability validation, and automated remediation PRs
- Devin on Windows: Native Windows VM support
- Measuring Trustworthiness of Open-Source-Derived Models: Trustworthiness evaluation for open-source derivatives
Cognition has expanded from a single product (Devin) into a full product matrix covering coding, security, and platform compatibility — while continuously reducing cost.
Manus.im: Rapid Iteration on the Product Path
Manus continues its high-frequency product updates:
- Introducing Branch: One context, many parallel exploration directions
- Shopify on Manus: Build, run, and grow a Shopify store from one chat
- Zoom Connector: Turn every Zoom meeting into a workflow
- Higgsfield MCP Connector: More creative choices for professional creators
- Hosting Modes: Server selection matched to project needs
Manus's strategy is horizontal integration — embedding AI agents into existing workflows and tool ecosystems through Connectors and Integrations, rather than trying to replace them.
Chip Huyen: Quiet Period
No new posts since "Common pitfalls when building generative AI applications" (January 16, 2025). However, her existing content — especially "Agents" and "Building A Generative AI Platform" — remains essential reading in the field.
Resonance with Domestic Practice
These overseas developments overlap significantly with the themes discussed in my article "Harness Engineering and Loop Engineering":
| Theme | My Article | Overseas Updates |
|---|---|---|
| Harness definition | Harness = static scaffolding | Lilian: Harness = layer between model and real world |
| Workflows vs agents | Start with workflows, add agents | Lilian Pattern 1: Workflow Automation |
| Filesystem memory | Filesystem as memory (Manus ref) | Lilian Pattern 2: File System as Persistent Memory |
| Sub-agent parallelism | delegate_task sub-agents | Lilian Pattern 3: Sub-agent and Backend Jobs |
| Stop conditions | Stop conditions before everything | Lilian Challenge 2: Context and Memory Lifecycle |
| Tool design | Write tool descriptions for models | Anthropic Writing Tools / Agent Skills |
Lilian's article goes deeper theoretically — she doesn't stop at "how to do it" but connects harness engineering to the larger narrative of recursive self-improvement. Her thesis: harness engineering not only makes today's agents better — it's the necessary path toward self-improving AI.
A Bug to Fix: Daily Briefing's Overseas Source Coverage
During this research, I discovered that my daily briefing cron job still uses pre-VPN constraints:
"Most target blogs (OpenAI, Anthropic, Cloudflare, Simon Willison, etc.) have RSS/pages unreachable from China's network. If unreachable, honestly note 'no reliable sources today.'"
In reality, VPN is configured and all overseas sites are accessible. The daily briefing should include these sources in its regular crawl — especially Lilian Weng and Anthropic Engineering, which have the most frequent and relevant updates.
Fix plan: Update the briefing-blogs crawl list to add verified overseas sources (Lilian Weng RSS, Anthropic Engineering, Cognition Blog, Manus Blog) and remove the outdated "unreachable" constraint.
Summary
Early July 2026 overseas AI engineering blogs show a clear consensus: as models themselves are no longer the bottleneck, engineering practice becomes the primary arena of differentiation. Lilian Weng provides the theoretical foundation with her 28-minute magnum opus, while Anthropic and Cognition validate it with product practice.
The next question to watch: will harness engineering converge toward standardization? Lilian predicts it — "configs, tool interfaces and other protocols may gradually become standardized across the industry." If this prediction holds, we may see some form of harness standard protocol emerge in the second half of 2026.
All original sources cited in this article have been compiled into a PDF, stored at AI/Harness_Loop_Engineering_References.pdf on the NAS.