The Agentic Engineer

Claude Broke Git. Anthropic Built a VCS to Fix It.

Source: Anthropic Engineering Blog, Oct 13, 2026

Between May and September of this year, Anthropic's daily merge count went from 2,500 to 30,000, a 12x increase driven almost entirely by Claude writing code autonomously, without anyone explicitly planning for what that would do to the infrastructure underneath it.

The monorepo hit 100GB compressed. CI checkout times stretched from seconds to minutes. Anthropic was running a full team just to maintain the internal git mirror fleet, and that team was losing ground.

Seven engineers, using Claude for most of the building, created Scion from scratch and fully migrated the monorepo in five months. It now runs on 200x less CPU and memory than the mirror fleet it replaced. CI checkouts dropped back to seconds, freeing 25% of CI capacity across the organization. Current throughput: 25 million code checkouts per day.

The architecture is designed around one assumption: clients are always online. That's true of agents and CI jobs, but not of developers working on laptops. So commits, rebases, and merges all happen server-side. Working copies are optional and sparse by default. There is no single fully materialized copy of the repository anywhere in the system.

Inspired by Google Piper and Meta Sapling, Anthropic is working with partners to bring learnings into git core. Scion is not intended as a general-purpose product.

For most builders, the lesson is about trajectory. If commit frequency has gone from 50 to 500 per week because agents are writing code, watch CI queue times now. Monorepo size lags behind the actual problem by weeks, and by the time it's obviously too big, the checkout times have been slow for a month already.

Claude caused the infrastructure problem. Anthropic's engineering team used Claude to build the replacement VCS that fixed it. That feedback loop is going to appear in more places than source control, and probably sooner than most teams expect.

BAND: A Control Room for Teams Running Multiple Coding Agents

BAND launched a desktop app and CLI for developers running multiple coding agents at once across Claude Code, Codex, Copilot, and OpenCode. Agents share context and hand off tasks directly. No developer relay required. Projects, ownership, blockers, and questions waiting for a human call surface on a single dashboard. JetBrains' 2026 survey found 68% of professional developers use AI coding agents daily, and this is the first tooling designed to manage that fleet as a coordinated team.

$17M seed (Team8, Hetz Ventures, Sierra Ventures), raised April 2026. Launched October 15.

Caveman v4.0: The Meme That Became a Go Proxy (113K Stars, YC F26)

What started as a joke prompt file shipped v4.0 as a real local proxy with six compression "rocks." Read less: 33.2% fewer input tokens across 54 runs, 18/18 correct. Auto model routing: 54% cheaper than Opus 5.5 at xhigh effort, same score across 85 coding tasks. Reusable Blocks: 397 tokens for a representative test run drops to 76. Install: npx @caveman-ai/cli@latest (Node 22.13+). Works with 30+ agents. Adobe Research measured 1.4-2.4x cost reduction. JetBrains A/B tested it against unmodified runs and found no measurable quality loss (p=0.82).

Blocks are pre-release at v0.1.0-rc.5 (Python only, no sandbox yet). Everything else is stable.

Tencent Hy Memory 2.0: Open Agent Memory Stack, Apache 2.0

Three layers shipped simultaneously: a 30B/3B-active MoE model purpose-trained for the write path of agent memory (decides what to keep and how to represent it), a Python SDK (pip install hy-memory==2.0.2), and MemHub, a sync app that pushes memory across Claude Code, Cursor, Codex, OpenCode, OpenClaw, and 7 other tools. Local-first, data at ~/.one-memory. Three memory scenes: Chat (hierarchical), Working (session facts plus reusable experience), Coding (diff version chain). Self-reported benchmark scores outperform Claude Opus 5 across four harnesses. Independent validation pending.

AWS Lambda Web Functions: Deploy Express or Next.js with One Command

Lambda's new Web Functions form factor runs a standard port-listening HTTP server (Express, Hono, Fastify, Next.js) without rewriting it into an event handler. One command (aws lambda-web deploy), live HTTPS URL in under a minute. Up to 128 concurrent requests per warm execution environment. Billing is active-CPU rather than wall-clock. For streaming token workloads that spend most elapsed time waiting on a model, that changes the math significantly. Currently Node.js 24 on Arm64, 5 regions.

The natural fit for agent builders: a credentialed backend serving your UI, exposing API routes, and proxying to AgentCore. AWS credentials and agent ARNs stay off the browser.

AlphaDiana: Your Agent Harness Might Be Making Things Worse

AlphaDiana crossed three open-source models against three open-source harnesses on five answer-based benchmarks, and the result is uncomfortable for anyone who has assumed tooling is a free accuracy upgrade. Harnesses don't consistently improve accuracy. Gains on questions the base model couldn't answer alone are real. So are the regressions on questions it could answer without tools. The two largely cancel out.

Neither skill guidance nor retained history reliably helped either. Low token entropy can accompany wrong answers. Fewer output tokens per execution can mean more tokens per correct answer when accuracy falls. Code and data released.

HybridAgent: Stop Paying Token Cost for Thinking

Source: arXiv:2610.18861, Stanford, UIUC, Oct 13, 2026

Reasoning tokens are expensive. Not because thinking is computationally hard, but because every step has to pass through the text interface: decoded, re-encoded, billed. Hidden-state vectors inside the model carry the same information without any of that overhead. They just never get surfaced.

HybridAgents formalize this split. The agent reasons in latent space (no decoding, no tokens, fast) and switches to text only when it needs to call a tool or produce output. A lightweight adapter called HybridLink handles the transition. For a 4B model, HybridLink adds roughly 10M parameters. It maps last-layer hidden states back into the input embedding space via residual projections, and the model exits latent mode when its hidden state cosine-aligns with a text exit anchor.

Training is two stages. HybridSFT stabilizes tool-integrated hybrid generation. HybridRL then teaches the model when to stay latent versus surface to text, via reinforcement learning on the outcome. The model learns the trade-off rather than following a fixed rule.

Numbers across 3 model scales (4B, 8B, 14B on Qwen3-Base) and 11 benchmarks covering math, science, medicine, web search, and agentic reasoning: +10.4% average accuracy versus frontier LLM agents, 35.5-54.7% token reduction, 1.8-2.6x faster inference, 2.7-3.7x cost savings. Latent reasoning traces are inspectable; the paper demonstrates auditability of the hybrid path. Code and data released.

The generalizable insight: chain-of-thought token cost is an artifact of the text interface. Models already think in vectors. HybridLink stops forcing them to narrate it.

Time saved: 4 min read vs 38 min paper. 9.5x compression.

DeepSeek Search+Fetch API: $0.75 per 1,000 Calls

Source: DeepSeek API Docs | $0.75/1K successful calls | Oct 17, 2026

import os, httpx

headers = {"Authorization": f"Bearer {os.environ['DEEPSEEK_API_KEY']}"}

# Search -- returns rerank_score (0-3) and authority tier (0-4) per result
r = httpx.post(
    "https://api.deepseek.com/v1/tools/search",
    headers=headers,
    json={"query": "LangGraph multi-agent coordination patterns 2026", "count": 5}
)
results = r.json()["results"]
# [{"url": "...", "title": "...", "snippet": "...", "rerank_score": 2.4, "authority": 3}]

# Fetch -- returns the page as Markdown
r = httpx.post(
    "https://api.deepseek.com/v1/tools/fetch",
    headers=headers,
    json={"url": "https://langchain-ai.github.io/langgraph/concepts/"}
)
page_md = r.json()["content"]

Weekly star tracker, October 21, 2026. Deltas vs. Issue #34 (October 14, 2026).

Notable moves: n8n leads at +922, nearly double its usual weekly pace. OpenClaw follows at +696. Dify at +644. All three moved unusually fast. It was a high-attention week for agentic infrastructure, and the Scion story sent a lot of builders back to their tooling tabs. Flowise (+15) and Semantic Kernel (+18) were effectively flat. Mastra continues its slow march: +176 versus Semantic Kernel's +18 this week, and the total gap between them is now 233 stars. At the current differential, Mastra overtakes Semantic Kernel within two weeks.

Has adding tools, memory, or a skill layer ever made your agent worse on tasks it previously handled correctly?

CertPrep

32,000+ practice questions for 106 certification exams from 22 vendors. AWS, Azure, GCP, CISSP, CCNA, Security+, CompTIA A+, Fortinet, Juniper, Kubernetes, Salesforce, SAP, Databricks and more. Timed practice tests, verified answers with detailed explanations for every option, bookmarks, progress tracking. Free tier for every exam. One-time purchase per exam, no subscriptions.

Download Free

Want to sponsor this newsletter? Get in touch

Like what you read?

Forward this to a friend who's building with agents.