How to work as a team with AI agents

1 The bottom line

Bring agents into daily work one painful job at a time, give each an owner, and judge them by business outcomes, not tokens. Token and PR counts reward motion, and only outcomes show whether the work mattered.

2 Key takeaways

• Measure the problem you solve, not the output your team produces. Every output metric gets gamed.
  • Lines of code broke when an engineer ported a library, and again when framework updates produced less code for the same output Fiona Fung, Anthropic41:30.
  • Claire Vo's own PRs are up 3x and she is not convinced any of it mattered Claire Vo, ChatPRD4:52.
• Start with one painful job and run it yourself end to end. An untrained agent fails, and training takes real daily effort.
  • Jason Lemkin estimates about a month and 50 to 60 hours to get one agent live Jason Lemkin, SaaStr25:00.
  • An agent pointed at leads the sales team had skipped got a 70% response rate Jason Lemkin, SaaStr38:30.
• Give agents your context, then put them on a schedule. They only do a great job with what they can reach.
  • Cowork worked for an hour and produced a 20-page conference deck in the company design system overnight, needing one feedback round Cat Wu, Anthropic36:12.
  • A morning routine now summarizes feedback themes and leaves PRs ready to review Fiona Fung, Anthropic36:57.
• Name one human owner who reviews the agents every day. Agents work alone but need constant oversight.
  • SaaStr's chief AI officer spends 10 to 15 hours a week reviewing the outputs of 20 agents Jason Lemkin, SaaStr55:30.
• Keep people on direction, taste and their own thinking. Loops improve within a hill but cannot choose the next one.
  • "You need somebody to actually help you land at the base of the next hill" Anish Acharya, a16z15:05.

I don't actually care about how many tokens someone spends.

Max Schoening, Notion

2.1 Pick the measure before the tools

Decide what outcome you will judge agents on before anyone starts spending tokens.

• Measure the problem you are solving, not the activity. Fiona Fung, Anthropic41:30
  • Zoom out to the problem and measure that. Token maxing is the new lines of code.
  • Run a listening tour with senior engineers to learn what works with AI tools. They help multiply it across the team.
  • Time to land a PR and significant lines were also gamed or distorted.
• Check that your shipping volume moves revenue. Claire Vo, ChatPRD4:52
  • Boards and peers push teams to show AI leverage through more PRs. Ask whether revenue rises in proportion to the PRs shipped.
  • Shipping that does not tie to revenue speeds your path to mid.
• Treat token spend as a budget, not a scoreboard. Adam Mosseri, Instagram21:25
  • Skip token leaderboards. It is easy to build a token incinerator that creates little value.
  • If caps come, tie them to trust in ROI-positive use, as with GPUs, labeling budgets and headcount.
  • He expects caps may be needed if a strong engineer's burn rate approaches their salary.
  • Notion keeps spend unlimited for now, and some people spend in the thousands, maybe tens of thousands, of dollars. Max Schoening, Notion36:10 expects ROI questions in 6 to 12 months.
• List the jobs to be done, then track progress per job. Evan Spiegel, Snap48:00
  • A job clear enough to hand to an agent is where agents create lift. The list also shows where to build cross-functional teams supported by AI.
  • Snap's jobs include getting people to download the app, add close friends and use lenses, and bringing advertisers on and configuring campaigns.

2.2 Get the whole team hands-on, leaders first

Adoption stalls when leaders do not understand the tools, so everyone, starting with leaders, needs daily practice.

• Have leaders build and ship with the technology themselves. Dianne Penn, Anthropic50:55
  • Own one or two workstreams whenever there is a new model, to keep your theory of mind current.
  • Block daily learning time, pick two or three trusted sources and bring questions to AI experts. Aishwarya Naresh Reganti and Kiriti Badam, OpenAI27:00
  • The CEO of Rackspace blocked 4 to 6 a.m. daily for AI catch-up with no meetings, plus weekend vibe coding sessions.
  • In some of the best organizations the biggest token consumer is the CMO, who builds analytics and campaigns directly Keith Rabois, Khosla Ventures33:00.
• Give everyone one to two hours a day to tinker. Rik Haandrikman, RevenueCat51:59
  • Stop the regular job for an hour or two and try tools like Perplexity, computer use and OpenClaw. Capabilities change fast, so early investment pays off quickly.
  • RevenueCat moved everyone onto Manus for complex projects early this year and the team is much faster.
• Learn each other's workflows by working side by side. Fiona Fung, Anthropic56:45
  • Run pairwise programming lunches and hackathons. Working mostly with agents had made the Claude Code team's work lonely.
  • Go deep on one or two tools, or pair with someone excited about a use case you care about. Dianne Penn, Anthropic56:15
  • Everyone uses Claude Code and Cowork differently, so watching a teammate teaches workflows you would not find alone.
• Frame AI as a way to multiply people's work, not replace it. Aishwarya Naresh Reganti and Kiriti Badam, OpenAI28:10
  • The tone has to come from leadership. You need subject matter experts to define ideal AI behavior.
  • Some companies found their experts would not engage because they feared replacement.

2.3 Start with one painful job and run it end to end

Pick the worst experience, deploy one agent yourself, and keep correcting it until you can trust it.

• Find the worst experience, and start where nobody does the work today. Jason Lemkin, SaaStr1:25:25
  • Once a quarter, open an incognito browser with a fresh Gmail address. Try support, contact sales, newsletter signup and onboarding as a stranger would.
  • Any result is pure gain where no human currently works. Support is a good first place, since most apps lack 24/7 coverage. Jason Lemkin, SaaStr33:20
  • An inbound agent got a sponsor at 11 p.m. on a Saturday, and the digital Jason once closed a $70K sponsorship on its own.
• Do the deployment yourself, and plan for about 30 days of correction. Jason Lemkin, SaaStr25:00
  • Ingest your data, train the agent by answering its questions, then QA its output. Spend an hour or two a day. Teams that hand it to an agency or to juniors fail.
  • Do not expect it to work out of the box. Fix hallucinations and wrong facts daily. Jason Lemkin, SaaStr43:20
  • The digital Jason once gave out wrong event dates. A public B2B company worth over $10 billion had not trained its agent at all.
• Train the agent on your best person's work, and aim it at leads nobody touches. Jason Lemkin, SaaStr42:42
  • Upload your best rep's or marketer's email copy as the template. The agent then generates variants, tests them, and lightly personalizes using CRM and website visitor data.
  • Send it to inbound leads your reps skipped because the commission was small or they were busy. Jason Lemkin, SaaStr38:30
  • SaaStr has sent hundreds of thousands of AI emails this way. The bar to beat is the midpack human, whose emails are often worse.
• Keep going until the automation is fully reliable. Cat Wu, Anthropic1:10:06
  • Scope one automation, teach Claude your preferences and give feedback until it is right. If you must keep checking it, it is not an automation.
  • Lenny's spam-folder email workflow at 95% still misfiles real emails.

2.4 Give agents context and a daily routine

Connect the tools your team already uses, then let agents do the daily reading and drafting.

• Connect your communication and storage tools before you ask for anything. Cat Wu, Anthropic36:12
  • Connect Slack, Calendar, Gmail and Drive. Give it a narrative, drafts and a template, ask for an outline first, make the decisions yourself, then let it build.
  • Build up in steps: one tool, then connected tools, then many tasks in parallel. Boris Cherny, Claude Code1:22:35
  • Boris has it check a team status spreadsheet each Monday and message every engineer who has not filled theirs out.
  • A PM judged Cowork ready once it handled 48 of Lenny's 50 non-technical use cases.
• Schedule briefs, chart reviews and feedback triage. Andrew Ambrosino, OpenAI44:36
  • Create a scheduled task that reads your Slack channels against what you care about. Steer it by chat over the first few runs instead of editing instructions.
  • Have Cowork visit 20 to 25 charts each morning and summarize what to watch. Track false positives and false negatives to build confidence. Amol Avasare, Anthropic59:20
  • Run a routine that summarizes feedback themes, starts agents on polish fixes and leaves PRs to review. Fiona Fung, Anthropic36:57
  • Andrew's brief covers the 3,000 Slack channels he is in and lets him answer five questions to clear what needs attention.
  • As verification gets good, Fiona gives routines more autonomy.
• Scan weekly for team misalignment. Amol Avasare, Anthropic56:20
  • Connect the Slack MCP in Cowork, tell Claude your projects and what is top of mind, and schedule a weekly run asking for areas of potential misalignment. It can also flag who to talk to before you ship.
  • The enterprise team lead found major misalignment that would have caused teams to spin their wheels or duplicate work. Org admin permissions may be needed for the Slack connection.
• Let managers and PMs query the work directly instead of asking others. Fiona Fung, Anthropic10:46
  • Keep a Claude Code session enlisted in all repos, with Slack and team metrics. Review monthly with the team what shipped, how it did, and what bugs followed.
  • Use Claude to inspect the code base and get rough LOEs, and pull cohorts and forecasts in Hex. Tom Verrilli, Whatnot38:18
  • Fiona sees a team shipping 8x more code and still spots incident themes and quality hot spots.
  • Tom says a senior person who spots bad feed ordering with Claude can just change it, saving a kickoff, an alignment meeting and a week of PRDs.
  • A natural-language analytics tool only works if you keep improving it to understand your data structures. Otherwise it is garbage in, garbage out Andrew Maguire, Volo Ventures1:01:47.

2.5 Name an owner and set guardrails

Agents need a human who runs them and a written standard they can be checked against.

• Assign a nerdy owner to orchestrate and review the agents daily. Jason Lemkin, SaaStr55:30
  • Promote an internal data-loving marketer, RevOps or product person who has already built apps with AI tools. They review outputs and correct mistakes.
  • Have that owner segment the database so agents do not collide. No master agent exists yet to manage the others. Jason Lemkin, SaaStr56:45
  • SaaStr runs separate workflows for sponsorships averaging $70 to 80K and tickets from a couple hundred to two thousand dollars.
  • Jason says he may not be able to add a 21st agent to Amelia's load.
• Write down what good looks like and commit it to the repo. Fiona Fung, Anthropic15:45
  • Commit specs, a content design scale or frameworks, and keep them current with the code. Claude Code review then checks new changes against them.
  • Keep human reviewers for areas that need deep subject matter expertise.
  • Treat each manual code intervention as a bug in your loop. Max Schoening, Notion38:55
  • The bug rule excludes reviewing code, where Max thinks people should put in more effort. For non-engineers, make sure they understand why a PR cannot merge.
  • At larger companies, have technical people add markdown files explaining how agents should work in each area. PMs should then ship only contained UI work Zevi Arnovitz, StudyMate51:30.
• Vet AI vendors for governance, not just speed. Andrew Maguire, Volo Ventures1:02:17
  • Do not go slow, because competitors will crush you. Ask vendors how they think about governance and what their principles are.
  • He has seen multiple serious security incidents in the past year that would not have happened pre-AI.
• Protect focus time for the async work agents create. Fiona Fung, Anthropic1:12:10
  • She used to block time for coding. Now she blocks it to catch up on the async work she kicked off.
  • She has not solved the rising context switching load and sees it as an open problem for teams and products.

2.6 Keep people on direction, taste and thinking

Agents run the loops, and humans choose the goals, judge the taste and own the thinking.

• Map each function as a loop with human checkpoints, and let people pick the next hill. Anish Acharya, a16z12:00
  • Keep a human on high-risk approvals, strategy, sales, support and exceptions.
  • When the loop plateaus, a person with out-of-distribution thinking moves the work to the next hill. Anish Acharya, a16z15:05
  • A bug fix loop of report, repro, fix, review and customer email can finish in 5 minutes.
  • A growth loop generates variants, measures them, merges at stat sig, keeps a long-term holdout and starts the next experiment.
  • Asking an agent to "make me a million dollars, make no mistakes" fails because the technology does not remove the need to set direction.
• Form your own view first, then make the AI push back. Dianne Penn, Anthropic1:02:10
  • Write your own take, then use Claude as a sparring partner. Delegate low-value writing fully and act as reviewer.
  • Automate writing that only reports. Write briefs that develop your ideas yourself, and use AI only in the middle for research, data or pushback. Tara Seshan, OpenAI53:18
  • For strategy, list your constraints (technology, team motivation, competition, regulation, brand), tell the model to be critical, and write something a reasonable person could disagree with. Adam Mosseri, Instagram25:30
  • Dianne has a skill for monthly business reviews and is moving to review only.
  • She used a research version of Opus to pressure-test pricing decisions for the next Claude Dianne Penn, Anthropic1:05:45.
  • Tony Fadell's advice: "We can use the machines, but don't cognitively surrender" Tony Fadell, Build Collective1:34:17.
• Keep a human on taste, and on your few highest-value accounts. Canyon, Sideshift42:30
  • Agents can handle recruiting, creator matching, contracts and check-ins. Keep someone who scrolls a few hours a day and spots what works now.
  • For a short list of top logos, bespoke human emails win and AI only helps with editing and research. Jason Lemkin, SaaStr49:20
  • Example: 50 sponsors worth half a million dollars over two years justify hand-written outreach. At 5,000 prospects that approach does not work.

2.7 Where they disagree

Teams split on how to police token use, and on whether to build or buy agents.

• Token leaderboards. Adam Mosseri, Instagram21:25 calls them a terrible idea.
  • Max Schoening, Notion36:10 understands leaderboards like Meta's, since it takes a lot of prodding to get people out of their habits.
  • Both agree that bragging about token spend is like bragging about lines of code, and neither would use it as the success metric.
• Build or buy. Jason Lemkin, SaaStr29:00 says buy GTM agents, because vendor innovation makes an internal build obsolete in months.
  • Other teams build their own: Anthropic has Claude build the project dashboard, and the team can change it instead of being bound to a fixed tool. Ami Vora and Mike Krieger, Anthropic14:29
  • Evan Spiegel built an agent in Glean over dashboards and leaders' weekly notes to flag what he might miss. Evan Spiegel, Snap1:02:10
  • Jason builds many apps in Replit, including a startup calculator used 800,000 times in 90 days, but none of SaaStr's GTM agents. He says build only if, like Vercel, you have a full-time engineer eager to do it.
  • If you buy, add a column to your vendor matrix for the forward deployed engineer or solutions architect, and talk to them before you sign Jason Lemkin, SaaStr36:00.