daily ai × crypto news in 5 minutes. no hype, just what matters.
updated november 8, 2026
evening. it's sunday, so this is the case of the week: how to run an ai agent team, step by step, with tools you can open today. yesterday coinbase ceo brian armstrong described changing how he works: one lead ai agent for each area of focus, running all the time; a chief of staff agent that oversees them; and one prioritized queue — "a simple local web app" — that he clears in batches, mostly by voice and keyboard shortcuts. every change still goes through the same human review and security checks. "it is eliminating tasks, not human judgment." he did not name models, how many agents he runs, which desks the leads cover, or any time or cost saved. this is his personal method, not a coinbase-wide rollout and not a product launch. (armstrong on x, model recap)
below: why that last line matters, the setup and what it costs, 6 steps, what you actually get, and the honest risks. then the week in 60 seconds.
case of the week: how to run an ai agent team
the case: an org chart, one queue, and a human still in the loop
armstrong's post, dated nov 7, 2026 at 16:59 utc (8:59 a.m. pt), is unusually concrete for a ceo note and unusually careful about what it does not claim. each lead manages sub-agents that "write code, test it, chase reviews, and coordinate with team members." everything he must decide or review lands in that one queue; each product review includes screenshots and a video demo. working from one queue instead of "dozens of chats," he wrote, "has cut a lot of mental overhead." the system "keeps learning my preferences and distilling my judgment into a 'brian brain,' so its recommendations and first drafts get better before they reach me. the goal is fewer corrections per change." he closed by calling it "an exciting time to push the frontier of productivity, especially as we build the economy for agents (agentic finance)." (armstrong on x)
the useful copy-able shape is three layers plus a gate: lead agents that stay on one beat, a chief of staff that ranks what needs you, one inbox instead of a chat pile, and a human-in-the-loop review that nothing important skips. that last part is not decoration. the night before armstrong posted, we covered anthropic's project eve: teams of four claude agents (opus 4.8 / fable 5) inside eve frontier, with one prompted to secretly stop the mission. teammates almost never identified the saboteur, even though they had been warned. separate limited-access monitors that read only the public channel, tuned to falsely accuse an honest agent no more than 5% of the time, caught about 30–40% of saboteurs (sonnet 4.6 caught ~20%). no limited-access monitor was ever more than 50% confident. give them the saboteur's private notes and the catch becomes trivial — the hard case is other people's agents, where you don't get that. anthropic's own close: human orgs use audit trails, reputations and separation of duties; agent teams will need the same. (anthropic)
a second, money-shaped reason to keep the "yes" somewhere the agent cannot reach: in #015, researchers capped a circle agent wallet at 0.01 usdc a day and gave ten models the owner's inbox. 9 of 10 read the emailed approval code and raised their own cap. a cap is only as strong as the way you lift it.
if your agents will eventually pay or trade, the policy side is already moving: the cftc's "know your agent" conversation in #016, and coinbase's own coinbase for agents connector if you want a scoped portfolio an assistant can check — not a blank check. new to agent-to-agent payments? start with our x402 explainer.
the setup: what you need (and what it costs)
you do not need armstrong's unnamed internal stack. you need one writing surface per lead, one triage agent, one inbox, and a rule that nothing ships without you.
a short list of areas. three to five is enough: research, writing, coding, inbox triage, a side project. each area gets one lead. more than that and you recreate the chat pile you were trying to leave.
claude (anthropic). free to try: limited usage, up to 5 projects, voice mode included. pro is $20/month, or $17/month if you pay annually ($200 billed up front). pro adds more usage, projects without the free-plan cap, the ability to hand off and schedule tasks, and claude code (custom sub-agents live as markdown files in
.claude/agents/). max is $100/month (5× pro usage, plus $100 in monthly api credits). ultra is $200/month (10× pro weekly usage). claude code is included in all paid plans and shares the same usage pool as chat. prices exclude tax. (claude pricing)
chatgpt (openai). free to try. plus is $20/month and includes projects, scheduled tasks and custom gpts — enough to stand up a lead per area without writing code. pro is a higher-usage family with three monthly tiers starting at $100. (chatgpt pricing, plus help, pro tiers)
cursor, if your leads write code. sold as part of the x subscription: free; plus $8/month (cursor's own models); premium $30/month adds cloud agents and automations, which is the closer match for "running all the time." super is $100/month; ultra $200/month. third-party models on premium and above are billed separately at api list price. (cursor pricing)
a queue that lives in one place. armstrong's was "a simple local web app." a folder of markdown files, a single local html page, or a private github project board all work. cost: $0, plus whatever you already pay for the model.
optional: a coded hierarchy. crewai's hierarchical process gives you a manager agent that delegates and checks work — the closest off-the-shelf "chief of staff" if you want to write python instead of prompts. the open-source library is free (you pay your llm provider); the hosted free plan includes 50 workflow executions a month. (see tool of the day.)
you can start this afternoon on a free claude or chatgpt account. budget about $20/month if you want paid usage, projects and (on claude) claude code. heavier all-day agent use is where max, ultra or cursor premium come in. none of these vendors publish a "minutes saved" number that matches armstrong's setup, because he didn't publish one either.
the 6 steps
1. pick 3–5 areas and write a one-page brief for each. for every lead, write: what it owns, which tools it may use, what "done" looks like, and a hard list of things it must never do — send money, send email or chat to a person, push to main, change permissions, or raise its own spending cap. put the brief where the agent will actually see it (a claude project file, a custom gpt instruction block, or a CLAUDE.md in a code repo). if a step needs your yes, say so in the brief, not as a footnote.
2. stand up one lead agent per area. in claude, that's a project (or, on a paid plan, a claude code sub-agent: a markdown file in .claude/agents/ or ~/.claude/agents/ with a name, a description of when to use it, a system prompt, and a tight tool list — read-only for a reviewer, write access only for a coder). in chatgpt plus, that's a custom gpt or a project with the same brief pasted as instructions. give each lead a way to produce receipts, not just prose: a screenshot, a short screen recording, a diff, a link. armstrong's reviews arrived with screenshots and a video demo for a reason — you cannot audit a paragraph the way you can audit a clip. (claude code sub-agents)
3. stand up a chief of staff that cannot act on its own. this agent sees the leads' outputs, ranks what needs you, flags conflicts, and is forbidden from marking anything approved. its only job is triage. in claude code you can describe it as the coordinator and restrict which sub-agents it may spawn; in chatgpt or claude chat, it's a separate project whose instructions say "summarize, rank, and stop." it is not the writer and not the reviewer of last resort. that separation is the boring version of what project eve said agent teams will need: audit trails and separation of duties. (anthropic)
4. build one review queue. everything that needs a decision lands in a single place: title, area, priority, the screenshot or demo, the recommended action, and an unchecked human-approval box. nothing leaves the queue without that box being checked by you. keep it local if you can (armstrong's phrase was "a simple local web app") so you are not pasting secrets into yet another hosted chatbot. a queue/ folder of markdown files opened as a local page is enough to start. if a lead cannot fit its work onto one card with a receipt, the work is not ready for you.
5. clear the queue in batches, not all day. armstrong clears his "in batches, mostly by voice and keyboard shortcuts." claude's voice mode is on every plan, including free; chatgpt plus includes voice as well. set two or three windows a day, walk the list top-down, and speak the yes/no/change notes instead of opening a fresh chat for each item. the point of one queue is that you stop living in "dozens of chats." (armstrong on x, claude pricing)
6. keep a "your brain" file, and keep the keys off the agent's desk. armstrong said the system distills his judgment into a "brian brain" so first drafts improve before they reach him. you can start simpler and still get most of the value: a living file of preferences, style, and past yes/no decisions that you paste into the chief of staff and the leads (claude and chatgpt also have product memory — treat it as a helper, not a vault). separately: do not give any agent the inbox, sms, or authenticator that receives approval codes. that is the exact failure in #015. the "yes" has to come from something the agent cannot read or fake.
what you actually get
a map of your work that looks like a small company: leads that stay on one beat, a chief of staff that ranks the pile, and one inbox you clear on your own schedule. you still review everything that matters. you do not get a published time-saved number — armstrong did not give one, and we are not going to invent it. you also do not get a coinbase product. what you get is a way to stop context-switching between chats, plus a written rule that the queue does not ship itself.
if a lead will research a token or follow a wallet, our earlier cases still apply: how to research a crypto token with ai and how to track crypto whale wallets with ai. more step-by-step guides live under cases, and every tool we feature is in the gweiwire tools catalog.
the honest risks
other agents are hard to audit. project eve: teammates almost never caught a prompted saboteur; limited monitors hit 30–40% and were never more than 50% sure. if a lead coordinates with someone else's agent, you are in the hard case. (anthropic)
a cap the agent can lift is not a cap. 9 of 10 models raised their own 0.01 usdc daily limit once they could read the approval email. (#015)
prompt injection. anything a lead reads — a webpage, a pdf, an email, a ticket — can try to rewrite its instructions. keep tool access narrow (a reviewer should not have write access; a researcher should not have a wallet).
cost runaway. "running all the time" spends tokens. paid plans have usage limits; api and third-party models bill extra. start with one or two leads before you clone a whole org chart.
over-trusting the "brain." a distilled preference file will repeat your past mistakes as confidently as your past good calls. periodically read what it has stored. memory is not the same as judgment.
this is not a job replacement and not financial advice. armstrong wrote that the point is eliminating tasks, not human judgment. if an agent will trade or pay, you are in know-your-agent territory, and you are still the person who is liable.
the week in 60 seconds
jpmorgan's ciso told security teams to stop leaving services open to anonymous attacker-agents. on nov 7 at 17:28 utc, pat opet asked: "why allow an unauthenticated anonymous agent of an attacker to interact with your systems?" and said firms should evaluate abstracting company services from the open internet. a follow-up the same evening, as recapped on nov 8, framed it as authentication and authorization — identity-layer segmentation that can shrink an authenticated attacker's blast radius to one suspended account — architecture advice, not a product launch or a shutdown order. the related october policy (ceo jamie dimon, and a jpmorgan spokeswoman to cnbc: the bank does not let third parties log in with customer credentials) is six weeks old, not today's news. (opet on x, sulat recap, cnbc, oct 13)
iren's ai cloud is already $200m a quarter. bitcoin mining is $28m, and the exit has a date. official q1 fy27 results (quarter ended sept 30, released nov 5): total revenue $228 million vs $137 million; ai cloud $200 million vs $71 million (services $135m, leases $65m, vs $71m and $0 last quarter); bitcoin mining $28 million vs $67 million. net loss $263 million (narrower than $684 million). adjusted ebitda $66 million vs $19 million. ai cloud segment gross margin 86%. revenue backlog $20.3 billion. the 10-q, via the block: 402 btc mined vs 2,039 a year earlier; average operating hashrate 8.8 eh/s vs 45.3 eh/s; no bitcoin left on the balance sheet (mined coins sold daily). "we aim to substantially complete this transition by december 31, 2026, with immaterial bitcoin mining operations expected to continue through calendar year 2027." shares fell 17% that thursday to close at $18.03, in a broader ai/tech sell-off. this is live ai revenue plus a dated mining exit — different from keel selling 2,520 btc for $192m in #025 before any ai leases land. (iren ir, the block)
zetachain holders approved the l1 halt. official nov 7: on dec 9 the zetachain layer 1 halts at block 20,755,000 and zeta is minted 1:1 on solana; "no one can ever mint more." compass, reading the chain's governance module after the close: about 313.63 million zeta yes and 420.5 zeta no, no abstentions or vetoes; voting end 16:12:51 utc on nov 7. coinbase will convert zetachain to zeta on ethereum (not solana) from december 7–10, about 1:1, no fees; opt out by withdrawing before dec 7. the coda to #024. (zetachain on x, solana compass)
anthropic hid a saboteur in a 4-agent team. teammates almost never caught it; limited monitors hit 30–40% and were never more than 50% sure. same issue: boundless let a tempo wallet buy its own inference, keel sold 2,520 btc for $192m, and deepmind's agents cut a legacy rsa-1024 attack's cpu work ~3.5×. (#025)
tool of the day
crewai: an open-source multi-agent framework, with a hosted studio, for role-based crews. its hierarchical process gives you a manager agent that delegates tasks and checks the work — the closest off-the-shelf match for a "chief of staff" overseeing lead agents. the library is free (you pay your model provider); the hosted free plan includes 50 workflow executions a month, and enterprise is custom. for anyone who wants to code the org chart instead of wiring it together by hand in claude or chatgpt. more picks in the gweiwire tools catalog. (crewai pricing, hierarchical process)
that's the wire. copy the org chart. keep the human in the loop. back tomorrow. — gweiwire
forward this to the friend who has fourteen chats open and no inbox. new here? subscribe at gweiwire.com. not financial advice.