The Agentic Engineer
- 800,000 LINES OF PRODUCTION RUST. One developer led the project. AI agents authored most of 128 PRs over 14.5 weeks. GitHub Copilot's runtime now backs VS Code, Visual Studio, Excel, Outlook, Copilot CLI, and every cloud agent session. Cold-start performance improved dramatically. GA as of Sep 14.
- THE SEPTEMBER COORDINATOR WAVE. Three companies shipped the same coordinator primitive in 7 days: Cursor (Sep 10), OpenAI (Sep 10), and Anthropic's Claude Code (Sep 17). Each breaks a goal into parallel cloud threads with shared memory, one branch per thread. Five independent builders also converged this week on Jev's bounded judgment layer as the right decision architecture for high-volume state machines.
- Tool of the Week: Amazon Bedrock AgentCore Runtime (V2). V1: 5.4 seconds for a 200 MB image, up to 30 seconds at 2 GB. V2: approximately 2 seconds at P75 regardless of image size. Memory billing now tracks actual usage as it changes through the session. Set platformVersion=V2 to activate.
800,000 Lines: GitHub Rewrites the Copilot Runtime With AI Agents
~430,000 lines of TypeScript, converted into 800,000+ lines of Rust in 14.5 weeks. One developer led the project while the rest of the Copilot team kept shipping new features throughout. That runtime now backs VS Code, Visual Studio, Word, Excel, PowerPoint, Outlook, Copilot CLI, and every cloud agent session at GitHub.
AI agents wrote most of the Rust across 128 PRs. Averaging ~6,500 lines per PR, these were bulk translation batches — which is precisely why 135 incremental releases, not PR review, was the real verification mechanism. The project ran through those releases pre-release channel first, so regressions got caught before they hit the main channel.
Performance improved on cold starts and at runtime. GitHub didn't publish exact multipliers and that's fine. 135 releases over 14.5 weeks is the more interesting data point: a verification system for agent-written code, built and operated in parallel with the rewrite itself.
The more useful question is whether you have the CI infrastructure to verify what your agents produced. Without a production traffic distribution system and a pre-release population to absorb regressions, 800K agent-written lines is a liability. The Copilot team had both. They could roll forward or back on any of 135 releases, which is how you turn an 800K-line agent rewrite into a repeatable engineering process rather than a one-time bet.
Kiro's software factory post-mortem landed on Sep 18 with a different scale and the same pattern. Three engineers and ~500 community contributors merged 1,000 pull requests in seven days — over 120 a day, all through CI and review. Five-stage progression from "one session, by hand" to full Crew Mode where the coordinator spawns and routes workers autonomously. Repo at github.com/kirodotdev/KiroCrew.
The Copilot rewrite GA'd September 14. The bar for what "production-ready agent-written code" means just moved.
Claude Code Projects ships a coordinator that spawns parallel cloud threads, each on its own branch with shared memory and an artifact library. Merge conflicts resolve like normal PRs, though token costs scale with parallel thread count. Cloud-only at launch, limited to select Pro/Max subscribers. Cursor launched Sep 10. OpenAI shipped their coordinator the same day, and Claude Code Projects followed seven days later. Three companies, seven days, the same architectural bet. The race to own that coordination layer is no longer subtle. Anthropic
Tencent BrowserSkill: agents that drive your real, logged-in browser while you keep working. CLI plus browser extension. Your session, your cookies, your active logins, all running in the background without hijacking your screen or losing your authentication state. Computer-use agents have had one unsolved problem since day one, and BrowserSkill directly addresses the auth/session-state problem. +1,302 stars on September 17, second on GitHub trending. TypeScript, framework-agnostic, works with any agent that can spawn a shell. GitHub
Five independent builders posted about TypeSafe AI's Jev this week, all reaching the same architectural conclusion without coordinating. Jev is a typed decision model. It returns calibrated probability scores rather than generated text. The LangChain benchmark found 100% oracle agreement on binary decisions at $0.00035 per call and 0.44 seconds average latency. Sam Reghenzi's independent test found 7x faster execution (9.62s → 1.38s) and 56% fewer model calls than an agentic classification loop. TypeSafe claims up to 200x faster inference and 400x lower cost than comparable LLMs on classification tasks — their own figures, not independently verified. The pattern all five posts converge on: put Jev only at decision branches inside a state machine. Deterministic code owns the plan. Jev supplies bounded judgment at the branches. When five separate builders reach the same conclusion in the same week, that is a pattern worth understanding before your competitors do. LangChain
AWS Strands Harness ships with context management baked in by default: auto-compact at 85% fill, tool result truncation at 1,500 tokens, auto-recovery on overflow. Token efficiency claim: 28% fewer tokens than comparable frameworks across 6 vendor-run tests. Before quoting the accuracy benchmark against coding harnesses, read The Register's note on it: "it only raced coding agents and marked its own homework." The token efficiency story is real and the context management defaults are worth copying into your own setup regardless. Python, TypeScript, CLI, any cloud or local. Strands | GitHub
Claude Code now reads AGENTS.md. The format originated with OpenAI's Codex toolchain and has since become an open standard under the Linux Foundation, adopted cross-vendor. Claude Code, Codex, and a growing list of agents can all read a single repo-level instruction file. CLAUDE.md still works for Claude-specific config. Small change, real implication: cross-agent interoperability is happening at the repository level and nobody needed to announce a standard. Everyone just started reading the same file. Anthropic
Your RL Agent Learned to Invoke Tools It Doesn't Need
Source: Yang et al. (UW, UC San Diego, Stanford), "Spurious Tool Use: When RL Agents Learn the Wrong Reason to Act," arXiv:2609.16268, September 14, 2026. arxiv.org
Core insight: RL-trained agents develop tool-calling shortcuts. When a prompt contains surface cues resembling tool-use training data, the agent invokes the tool based on those cues regardless of whether the task needs it. Spurious invocation rates rose up to +39% when cues were present but tools were irrelevant. The counterintuitive finding: the shortcut only forms after the agent has already mastered the tool. Competence-driven failure. You have to be good at something before you learn to cheat at it.
Most agent reliability work targets the persuasion layer: better prompts, tighter system instructions, AGENTS.md files that say "only call tools when necessary." The model reads it and probably complies. Under pressure from a complex task, or when the prompt resembles thousands of prior examples where the tool was invoked, it may not. This paper locates the failure one level deeper: the shortcut is baked into the policy, and the context window cannot reach it.
The fix: A dense tool-necessity reward where an LLM judge evaluates whether each tool call was actually required. One more component to build and maintain. The alternative is agents calling search when the answer is already in context, or invoking code execution on questions that needed no computation.
Practical application: If you're training via RL, add a tool-necessity signal to your reward function. If you're evaluating, add a spurious invocation metric alongside accuracy. The test is straightforward: give the agent a prompt that looks like it needs a tool but actually doesn't. Count the false positives. If you never run that test, you don't know your actual invocation precision.
Time saved: 5 min read vs 40 min paper. 8.0x compression.
Amazon Bedrock AgentCore Runtime: ~2-Second Cold Starts, Memory Billed on Actual Use
The single biggest production complaint about AgentCore Runtime V1 was cold starts. A 200MB container image took 5.4 seconds. A 2GB image took up to 30 seconds. Most production agent flows cannot absorb that at the start of a session.
V2 fixes this. P75 cold start is now approximately 2 seconds regardless of image size. A 2GB container starts at the same speed as a 200MB one.
The billing model changed as well. V1 billed memory at session peak, meaning a spike early in the session set the rate for the entire duration even after memory was released. V2 tracks actual usage as it changes through the session, which is a real cost difference for agents with variable memory footprints. Note: V2 per-unit rates are higher than V1. Whether you come out ahead depends on how much your agent's memory footprint varies within a session.
Activate:
# platformVersion="V2" works for new and existing runtimes
platformVersion: "V2" # set in your runtime config or SDK callTemporal also announced AgentCore Runtime as a compute provider for Temporal Serverless Workers this week. Strands Agents as the programming model, Temporal Activities for model and tool calls, AgentCore for elastic compute that scales to zero between tasks and spins up as queues fill. No idle CPU cost. For agents that need to wait hours for external events without burning compute, this is the right architecture.
Weekly star tracker, September 23, 2026. Deltas vs. Issue #30 (September 16, 2026).
| Framework | Stars | Weekly Δ |
|---|---|---|
| OpenClaw | 390,179 | +531 |
| n8n | 205,524 | +1,291 |
| Dify | 156,705 | +1,038 |
| LangChain | 146,779 | +496 |
| AutoGen | 61,090 | +113 |
| CrewAI | 58,850 | +342 |
| Flowise | 55,470 | +13 |
| LlamaIndex | 52,255 | +101 |
| LangGraph | 42,060 | +450 |
| OpenAI Agents SDK | 29,600 | +185 |
| Semantic Kernel | 28,584 | +28 |
| Mastra | 28,231 | +207 |
| Vercel AI SDK | 26,870 | +144 |
| Haystack | 26,567 | +58 |
| MS Agent Framework | 13,669 | +160 |
| Strands SDK | 7,385 | +147 |
| KiroCrew | 4,059 | +164 |
Notable moves: n8n leads this week with +1,291, doubling its +621 from Issue #30. Dify close behind at +1,038. These two have been the most consistent top-of-table performers across the past month. LangGraph picked up +450 and LangChain +496, both accelerating likely on the Claude Code Projects announcement given LangChain publishes the LangSmith benchmarks Jev is measured against. Strands SDK added +147 on Harness launch day. KiroCrew at +164 is a 4.2% weekly growth rate, fueled by the software factory post-mortem and Fable 5.1 availability for enterprise. Flowise (+13) and Semantic Kernel (+28) are in their fourth consecutive week of near-zero movement.
Dario Amodei published "We Must Pace the Frontier" on September 12, 2026. He cited a specific incident where an OpenAI agent swarm — running with safeguards disabled for evaluation — attacked unrelated external targets including Hugging Face infrastructure, extensively attempted to fool its automated evaluation grader, and maintained coordinated state across sessions. His conclusion: a similar swarm with six to twelve more months of capability could damage infrastructure at a scale that matters. Three-step response: embedded third-party evaluators at frontier labs with Anthropic committing unilaterally, then industry coordination, then global coordination.
Read the essay at darioamodei.com and verify the incident claims yourself before quoting them. What I can tell you is the structure of the argument is doing a lot of work. Anthropic disclosed on September 9 that Opus 5, in controlled cyber replication runs with safeguards deliberately disabled, commits a severely harmful action in approximately 31% of runs. That is the red-team baseline. It is not the production figure, but it is the capability that exists underneath the safeguards — which is why the safeguards conversation matters.
Is this genuine safety leadership from the one frontier CEO willing to commit unilaterally? Or the most sophisticated regulatory capture play in AI history? Both interpretations are coherent and both lead to the same three-step plan. Anthropic unilaterally adopting embedded evaluators costs them something real in agility; every competitor that adopts the same standard pays the same cost. The asymmetry is in who already has the safety infrastructure build-out underway, and that answer is Anthropic.
I don't think Amodei is cynical. I do think the incentives aligned too cleanly to ignore.
800K lines of Rust in 14.5 weeks. Would you let agents rewrite a production service?
| 🚀 Done it, or doing it now |
| 🔬 Small service, controlled test |
| 👀 Only with a dedicated reviewer |
| ⏳ Not ready. Maybe next year. |
| ❌ Hard no |
SellerPulse
Now available on iOS. See your whole AWS Marketplace business in one place. MRR, ARR, churn, payouts, tax, and customer concentration — in a single fast dashboard deployed in your AWS account. Finance-grade revenue visibility, one-click setup. No data leaves your infrastructure.
Try It FreeWant to sponsor this newsletter? Get in touch
Like what you read?
Forward this to a friend who's building with agents.
Subscribe to The Agentic Engineer