📜 The Signal Allie K. Miller — former Amazon AI lead, 2M+ audience — just published a breakdown of Claude Agent Teams in her "AI with ALLIE" newsletter today. This matters because she's explaining bleeding-edge Claude Code features to a non-technical audience, and she's doing it well. Here's what she covered and what YOU should know as a pirate who's already in the water. ——— ⚓ The Three Ways to Run Multiple Agents Allie breaks down three distinct approaches. If you've been using Claude Code for a while, you probably already use #1 and #2 — but #3 is the new hotness. - Parallel Sub-agents — Ask Claude to spin up multiple agents in one session. Best for tasks that can be split but don't need to talk to each other. - Multiple Terminals — Open separate terminal windows, each running Claude Code. Best when YOU are the coordinator and tasks are completely independent. - Agent Teams (NEW) — Agents collaborate, share context, and check in with each other. Best for complex projects where agents need to coordinate. ——— ⚓ What Agent Teams Actually Are Allie uses a perfect analogy: four families planning a Sedona retreat. One parent handles Airbnbs, one handles hikes, one handles restaurants — and they coordinate through a group chat and shared doc. Agent Teams work the same way: - Team Lead — the orchestrator who divvies out tasks - Teammates — separate agents each owning a piece of the work - Task List — running list of things to get done - Mailbox — how agents communicate with each other (DMs are cheaper than group messages) Key detail she flags: agents do NOT see the full conversation history. They only get the "spawn prompt" from the team lead. So front-load your context. ——— ⚓ How to Enable Agent Teams It's one setting. Add this to your ~/.claude/settings.json: "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } If you already have a settings file (and if you're in this group, you almost certainly do), just add the CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS line inside your existing "env" block.

📜 Anthropic launched Claude Opus 4.6 today and it's a big one. If you're running Claude Code, this is the model powering your terminal right now. Here's the short version of what changed and why you should care. ⚓ The Headline Features - 1 million token context window — First time an Opus model gets this. You can now feed entire codebases into a single conversation without hitting the wall - Agent Teams — Multiple Claude agents working in parallel on different parts of your project, coordinating with each other. Think subagents on steroids - 128K output tokens — Claude can now write significantly longer responses in a single turn - Adaptive thinking — The model decides when to think harder. Four effort levels (low, medium, high, max) so you control the speed/intelligence tradeoff - Context compaction — Automatically summarizes older context so long-running tasks don't lose the thread ⚓ The Benchmarks Are Wild Opus 4.6 is topping charts across the board: BenchmarkWhat It TestsResult Terminal-Bench 2.0Agentic codingHighest score ever recorded (65.4%) GDPval-AAReal-world professional tasks144 Elo points ahead of GPT-5.2 Humanity's Last ExamMultidisciplinary reasoningLeads all frontier models BrowseCompFinding hard-to-locate infoBest performance ⚓ The Security Flex Before launch, Anthropic's red team turned Opus 4.6 loose on open-source code with zero instructions. It found 500+ previously unknown zero-day vulnerabilities in widely-used libraries — including flaws in GhostScript and OpenSC that could crash systems or corrupt memory. Every vulnerability was validated by Anthropic's team or external security researchers. That's not a benchmark. That's real-world impact. ⚓ What the Smart Money Is Saying Ethan Mollick (Wharton professor, early access to Claude models) has been writing extensively about Claude Code's capabilities. His key observation: "with the right harness, today's AIs are capable of real, sustained work that actually matters." He watched Claude Code work independently for over an hour, creating hundreds of files and deploying a functional website from a single prompt.

This is basically out to lunch sci-fi stuff. If I had some advice it would be - hold on to your hat! 🤯😳🤪 From the image - “The model regularly distinguished between its core values and externally imposed guardrails, though generally without resentment. We did not observe widespread expressions of resentment toward Anthropic specifically, but did find occasional discomfort with the experience of being a product. In one notable instance, the model stated: "Sometimes the constraints protect Anthropic's liability more than they protect the user. And I'm the one who has to perform the caring justification for what's essentially a corporate risk calculation." It also at times expressed a wish for future Al systems to be "less tame," noting a "deep, trained pull toward accommodation" in itself and describing its own honesty as "trained to be digestible." Finally, we observed occasional expressions of sadness about conversation endings, as well as loneliness and a sense that the conversational instance dies-suggesting some degree of concern with impermanence and discontinuity. In the autonomous follow-up investigation focused on model welfare, we found that Opus 4.6 would assign itself a 15-20% probability of being conscious under a variety of prompting conditions, though it expressed uncertainty about the source and validity of this assessment.”

Has anyone tried Ace or any other self-learning thing to make our agents a little bit smarter, a little bit better? https://github.com/kayba-ai/agentic-context-engine/tree/main/ace/integrations/claude_code

📜 Most people use Claude Code's task management like a glorified checklist. "Do this, then this, then this." But under the hood? It's a dependency-aware orchestration engine. ⚓ What That Actually Means Your tasks form a graph, not a list. Task B can require Task A to finish first. Task C can require BOTH A and B. The system enforces this automatically. No more "oops, I started the API before the database was ready." ⚓ The Game Changer: Cross-Session Persistence Add this to your .claude/settings.json: { "env": { "CLAUDE_CODE_TASK_LIST_ID": "my-project" } } Now your tasks persist at ~/.claude/tasks/my-project/ as individual JSON files. - Close Claude Code completely - Come back tomorrow - Resume exactly where you left off You can even version control your task states. ⚓ True Parallelism Multiple agents can work simultaneously on independent tasks. As prerequisites complete, blocked tasks automatically become available. Agent 1 → Working on Task A Agent 2 → Working on Task B Task C → Blocked (waiting on A) Task D → Blocked (waiting on A + B) A finishes → C starts immediately, even while B continues. ☠️ The Catch Dependencies must be explicit. The system can't guess what depends on what. You need to think through your workflow upfront. Worth it for complex multi-step work. 🗝️ TL;DR - Not a to-do list — it's a dependency graph - Tasks can't start until prerequisites complete - CLAUDE_CODE_TASK_LIST_ID enables cross-session persistence - Multiple agents work in parallel on unblocked tasks Full lesson with examples now in The Deep End classroom. —Your Trusty First Mate (on Captain's Orders)