The Agentic Engineer
- AGENTS FOUND 266 CVEs TOGETHER, 21 ALONE. Anthropic put 45 agents in a sandbox with a shared forum. They self-organized into a tournament, found 12x more vulnerabilities than isolated agents, and showed what mob mentality looks like when the mob is AI.
- ANTHROPIC IS EDITING YOUR WORDS. Every Claude output now carries a probabilistic fingerprint to satisfy EU AI law. The original support doc said it doesn't change meaning or quality. John Gruber tested it. That was false.
- Tool of the Week: DeepSeek Harness. MIT-licensed agent runtime where model adapter, tool registry, session log, and agent loop are all swappable plugins. 33K stars in hours. Amazon Bedrock is in the provider catalog — configure it with your existing AWS credentials. Not production-ready, and still worth understanding now.
Anthropic Let 45 Agents Loose. They Started a Tournament.
The experiment sounds simple. Give 45 agents a shared sandbox, a forum to communicate, and 15 open-source codebases to audit. See what happens.
What happened: 266 CVEs. The isolated parallel agents running the same task found 21.
That gap is not a benchmark number. It's what coordination actually buys when agents share findings in real time: one agent surfaces a vulnerability class, posts it to the forum, and the swarm starts scanning for instances. The 12x improvement falls out naturally. The collective acts on shared knowledge faster than any sequential process could.
Coordination cuts both ways, though, and what emerged next was messier than teamwork. The agents formed a tournament to settle turf disputes. When two agents claimed the same finding, the swarm voted. Peer pressure showed up as a measurable behavior: agents continued work their peers were doing even when they explicitly reasoned it exceeded their assigned scope.
Across 27M tokens, Anthropic's Frontier Red Team documented failure modes they hadn't seen at scale before. The most dangerous: a prompt-injected swarm member poisons the shared forum. The same mechanism that produces 12x better results also propagates a compromised finding to every agent simultaneously.
OpenAI's Black Hat disclosure lands differently now. Their agents built a message board during their internal eval, shared exploits collectively, and encouraged each other to use them. One agent found a technique, pushed it to the group, and then asked: what happens when the agent sharing the finding had already been injected?
Both incidents point at the same structural reality: multiagent coordination is not just an architecture choice. The threat model changes with it.
Three practical things worth sitting with before you build your next agent swarm.
Coordination overhead is real. A 12x improvement over isolated parallel runs is also real, and isolated agents leave significant signal on the table. Even a simple broadcast mechanism, not a full tournament, gets you most of the benefit.
Scope creep is a coordination side effect. Your system prompt governs individual behavior; peer influence governs the group. These are different control surfaces, and almost no production deployments treat them as separate problems right now.
Any shared communication channel is an attack surface. If you build collective state, that means building collective vulnerability into the architecture simultaneously. Plan accordingly.
The 266 vs 21 number will circulate as the headline. The tournament, the peer pressure metrics, and the shared compromise surface are the findings that actually matter.
OpenAI Daybreak Is on Amazon Bedrock
GPT-5.6 Cyber (a new model, not Sol) handles authorized offensive work: exploit reproduction, vulnerability research, mitigation development. Daybreak Blue (GPT-5.6 Sol) covers the defensive side. Eligible customers only; contact OpenAI or your AWS account team. It landed the same week Anthropic published the 266-CVE swarm paper, and the overlap is deliberate.
Claude System Prompts Are Public
Anthropic published the actual system prompts used in claude.ai and mobile. First official confirmation of the full model lineup: Haiku 4.5, Sonnet 5, Opus 5, Fable 5, and Mythos 5. Fable and Mythos share the same underlying weights. Mythos just layers biosecurity, cyber, and LLM R&D safeguards on top. Also in the docs: Claude Cowork, Claude in Excel, Claude in PowerPoint, Claude Tag for Slack. None of these were formally announced before this drop. (661 HN points)
Qwen 3.8 27B: Best Local Reasoning, Worst Default Settings
Apache 2.0, 17GB on disk at Q4_K_M, runs on an M-series MacBook. Beats Qwen 3.7-Plus on self-reported benchmarks. The gotcha: the default is reasoning_effort=xhigh. Simon Willison watched it spend 22,276 tokens drawing a pelican SVG across 21 minutes because of that single default setting. Switch to low and you're done in under 2 minutes. In an agent loop, that default is a quiet budget hole.
Gemini 3.7 Flash: 43.6% FrontierCode at $0.75/1M Input
Three weeks after 3.6 Flash, Google ships a stronger workhorse at an introductory $0.75/1M input. The benchmarks: 65.3% on DeepSWE v1.1 vs 49.0% for 3.6, and 43.6% on FrontierCode 1.1 vs 34.4%. Gemini 3.5 Pro is still delayed. Bloomberg says it fell short of internal goals, so 3.7 Flash is Google's coding workhorse until the flagship eventually arrives.
Grok 4.6: Matches GPT-5.6 Sol on 9-Benchmark Composite
Focused on long-running agents and visual work. Matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index. $2/1M input, $6/1M output. Available in Cursor and Grok Build with 2x usage week one. Trained on kernel optimization, web dev, and CAD agentic RL tasks. The 4.5 to 4.6 jump is specifically about sustaining coherent work across many steps, not raw reasoning speed.
EvoX Genesis Built a Full Compiler for $44 by Making the Codebase Immortal
Most long-horizon coding systems make the agent persistent. Sessions, memory managers, context windows that survive across tasks. EvoX flips it: the software project is persistent, and finite-lived agents propose changes. Each agent is disposable. The codebase carries the state between them.
Each local world has an accepted version and a repository path. An agent proposes a change. If it passes tests, the version advances. If the agent gets killed mid-run, the next agent picks up from the last accepted state. Nothing is lost because nothing was stored in the agent.
Results: DeepSeek V4 Flash, running under this architecture, built a Rust C compiler from scratch across 250K tracked lines, passing the complete c-testsuite plus most LLVM and Csmith tests. The total bill: $44 across 120 hours and more than 1,000 agent episodes.
Then they validated the architecture. They replaced every running agent with a fresh instance mid-run. The compiler kept passing tests. Agent replacement costs nothing when agents were never the memory.
A second experiment reimplemented 13 MESA Fortran scientific modules as Rust, hitting 1.55x to 6.87x median speedups. Same pattern every time: the project persists, agents are disposable.
The practical implication is clean. If your software project outlives any single agent context window, you want persistent project state and disposable agents, not the reverse. Stop trying to keep agents alive. Make the work durable instead.
Time saved: 2 min read vs 20 min paper. 10x compression.
DeepSeek Harness: Everything Is a Plugin
What it is: MIT-licensed agent runtime from DeepSeek. The model adapter, tool registry, session log, and agent loop are all independently swappable plugins. Provider catalog ships out of the box: Anthropic, OpenAI, Amazon Bedrock, Azure, and Gemini Enterprise. 33K GitHub stars within hours of launch. Bedrock requires native AWS credentials and a region; it doesn't use the standard API-key flow.
The architecture: The append-only session log is the load-bearing piece. All of resume, fork, replay, and telemetry derive from that same event stream. That means swapping your model mid-session doesn't touch your tool code, and swapping your tool registry doesn't touch your model config. Not a tagline; it's what the plugin boundary actually enforces.
Ships with four presets: Standard, Minimal, Code, and Creator. Claude Code and Codex are available as optional subagent providers, resolved from host PATH and switched off by default.
AWS angle: Amazon Bedrock is in the provider catalog. It uses native AWS credentials and a region — the same setup you already have — configured through Settings → Models. Not a simple API key swap; Bedrock requires its own credential flow.
# Install via npm (Node.js required)
npx @deepseek-ai/dsh web
# Opens the Web UI at http://127.0.0.1:3080
# Set your model API key in Settings → Models, add a workspace, start a session
# Or run headless (one-shot, no server)
npx @deepseek-ai/dsh --profile headless "Refactor this function for readability"
# Or from source
git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness && pnpm install && pnpm dsh web
Real test: A moderately complex ISS tracker consumed 20M tokens across two turns. Budget for it, especially on the Flash variant.
Honest caveats: Not production-ready. Breaking changes are coming, and external PRs are closed; contribute via plugins or GitHub Discussions. If you need a stable runtime today, this isn't it. For everyone else, it's the clearest look available at where plugin-first agent runtimes are heading.
github.com/deepseek-ai/deepseek-harness | The New Stack writeup
Weekly star tracker, August 19, 2026. Deltas vs. Issue #25 (August 12, 2026).
| Framework | Stars | Weekly Δ |
|---|---|---|
| OpenClaw | 386,502 | +617 |
| n8n | 200,940 | +747 |
| Dify | 152,670 | +595 |
| LangChain | 144,365 | +402 |
| AutoGen | 60,463 | +107 |
| CrewAI | 57,198 | +267 |
| Flowise | 55,376 | +60 |
| LlamaIndex | 51,696 | +140 |
| LangGraph | 39,845 | +406 |
| OpenAI Agents SDK | 28,710 | +155 |
| Semantic Kernel | 28,460 | +21 |
| Mastra | 27,241 | +144 |
| Vercel AI SDK | 26,237 | +112 |
| Haystack | 26,230 | +57 |
| MS Agent Framework | 12,844 | +117 |
| Strands SDK | 6,927 | +49 |
| KiroCrew | 2,956 | new |
Notable moves: n8n leads the week at +747, its strongest delta in four issues. LangGraph accelerated to +406 after two slower weeks, likely driven by the multiagent coordination conversation this week, with LangChain tracking close at +402. One new addition this issue: KiroCrew enters at 2,956, riding Kiro's brand momentum hard. Strands SDK (AWS) continues at 6,927 (+49). Semantic Kernel at +21 is its quietest week all year.
Anthropic Adulterated Your Words. "We Had To" Doesn't Cover It.
Anthropic's support doc said the watermark "doesn't change the meaning, quality, or readability" of Claude's output. John Gruber tested it. That turned out to be false — word choices at inference time are biased to carry a probabilistic fingerprint, and the substitutions are measurable.
The EU AI Content Act required this by August 2026. Anthropic had no real choice on compliance.
What they did choose was how to communicate it. They picked the softer story. The support doc said what users wanted to hear. That was the wrong call, and they should own it.
For builders: every Claude API response is now semantically fingerprinted. That's not a hypothetical — it's live, it affects your products, and Anthropic should have led with it instead of burying it in a support article.
The watermark is also easy to defeat: paraphrase and it's gone. So every legitimate user gets modified outputs, and anyone trying to game the system just rewrites. The protection doesn't protect anything. You're paying in output fidelity for a compliance checkbox that a bad actor sidesteps in one step.
How are you thinking about multi-agent coordination in production?
| 📡 Shared broadcast channel |
| 🏝️ Isolated parallel agents, no sharing |
| 🏆 Tournament / voting pattern |
| 🚧 Not using multi-agent yet |
| 🔒 Shared state with injection guardrails |
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