Anthropic's Playbook for Running Human-Agent Teams
Anthropic published how its own teams actually work with Claude agents — and the four lessons are management skills, not prompting tricks. Here's the playbook, plus how I'd apply it as an EM without a fleet of agents.
- #agents
- #anthropic
- #management
- #delegation
On June 24, Anthropic published a look at how its own teams work day-to-day with Claude agents — and the surprise is that there's almost no AI technique in it. All four lessons are things a good engineering manager already does: write things down, define roles, set a north star, delegate in proportion to trust. The company building the agents is telling you the bottleneck isn't prompting. It's management.
That's great news if you're a Brazilian dev aiming at international teams, because management skills travel better than any framework. Here's the playbook, and what I'd actually do with it.
Lesson 1: if it isn't written down, the agent can't see it
Agents only know what exists as accessible text — docs, code, meeting notes, public Slack channels. Anthropic's teams deliberately default to public channels and write notes so agents can find them later; the payoff they report is agents that stop suggesting already-killed projects and start surfacing relevant work from other teams. Most "the AI doesn't get it" complaints I hear are really "we never wrote it down". Fix the documentation and the agent gets smarter for free — and so does your next human hire.
Lesson 2: agents get roles, not random tasks
Their engineering teams keep an explicit roster listing humans and agents, each owning distinct jobs — triaging feedback, reviewing code, reporting status — with tool access matched to the job (the QA agent gets Playwright, the data agent gets BigQuery). That's an org chart, not a prompt. The anti-pattern it kills: five devs each running their own private "shadow" assistant doing overlapping half-jobs nobody verifies.
Lesson 3: a written north star makes agents proactive
One internal team wrote down the goal "make product onboarding more helpful" — and an agent, unprompted, proposed rewriting the error messages in the onboarding flow. Onboarding success measurably improved the next week. Without a documented goal, an agent can only react to tickets; with one, it can suggest work you didn't think to ask for. Same as a junior engineer, honestly.
Lesson 4: autonomy is earned per task type
Anthropic reviews agent work closely at first, then widens scope only where the agent has a track record. Two mechanisms worth stealing: the Doer-Verifier harness — a second agent with clean context reviews the first agent's output, and precisely because it never saw the reasoning that produced the bug, it catches what the author-agent can't — and a weekly report where agents compile their own "lessons & missteps". One team went from reviewing every single decision to dispatching agents on 500 bug fixes independently. They didn't start there; they climbed there.
The career angle nobody says out loud
If work is becoming humans setting strategy while agents execute, then the skill being priced up is running a team — and you can practice it today without a single direct report. A dev who keeps a roster, writes context docs, and runs a Doer-Verifier loop is doing EM work. In an international interview, "I managed a team of agents with verification gates and earned-autonomy levels" is a story almost nobody else has yet.
Today: pick one recurring task you already trust an agent with halfway, and write its "role card" — the goal it serves, the context files it reads, what it may do alone, and what a second agent (or you) verifies. That one page is the seed of the whole playbook.
Sources
Keep reading
More guides like this one.
Split Your Agents by Context, Not by Job Title
Most multi-agent setups fail because people split the work the way they would split it across a team — planner, coder, tester. Anthropic's own guidance calls that the anti-pattern. Here is the split that works, and the bill you are signing up for.
Subagents: Spend Tokens Without Spending Your Context
Anthropic's own data says the tokens an agent spends predict most of its performance — but in one session, every token lands in the same window as your plan. Subagents break that link. Here's the three-line file, the defaults that changed, and when to skip them.
Skill, Workflow, or Agent? One Question Decides
Most devs reach for an agent when a 20-line workflow would be cheaper, faster, and deterministic. Here's the three-tier ladder with Anthropic's own definitions — and the single question that tells you which one to build.