Companies are building agent teams on demand
- Companies are connecting agents to shared knowledge and assigning them work as it arises.
- The new workflows still need evidence, risk-based checks and a person accountable for the result.
AI at work is starting to look less like a digital coworker with a permanent job and more like a team assembled for a specific task. On This Week in AI, Alex Finn said Grokbot’s tools and workflow made it more useful to his team for everyday knowledge work than Claude. The distinction matters: a model’s ability is only part of the package; the software connecting it to company tools can determine whether it gets work done.
The AI Daily Brief’s Nathaniel Whittemore says companies building around agents are investing in shared context: access to the information and decisions agents need to do work. Ramp brings Slack, Notion, Linear and Calendar into persistent memory, while Uber has built a context graph spanning more than 30 internal systems. Whittemore warns that stale information can undermine an agent’s work.
Instead of waiting for an employee to write a prompt, some systems start from a business signal. RunLayer can route tasks from customer calls, Slack messages or failed builds; Uber agents can begin code reviews and on-call triage, and trigger other agents. Anthropic’s research system assigns narrow tasks to sub-agents, with their number growing to fit the job. The company says this approach performed about 90% better than a single agent on its internal research evaluations.
That flexibility shifts the challenge from staffing a fixed roster of digital teammates to checking what each temporary team produces. Shopify requires security findings to come with a test demonstrating the bug; Stripe requires a person to review every code merge. Other companies let low-risk changes proceed automatically while reserving human approval for higher-risk work. Uber traces agent handoffs to a responsible person, keeping ownership attached to the outcome.
There is a reason to be careful about the teammate framing. Whittemore cited a Harvard Business Review study in which managers checked documents less carefully when they were labeled as coming from an AI teammate named Alex3. Among companies that already treated AI as employees, scrutiny fell about 16%. The emerging model may put agents in the workflow, but the people responsible for their work still need to look closely.
Source Intelligence
- Deep dive into what was said in the episodes
Meta’s Muse: How we launched the most successful consumer product this year | Mona Sarantakos • Nov 15
- Mona Sarantakos says she had worked on Meta AI for about two years when Nat Friedman urged a small group to try OpenClaw in late January or early February. After a weekend connecting it to email, calendars, and home automation, a brief prompt produced a useful weekly family summary.
- Mona says designer Alex Cornell's temporary Jolly avatar survived repeated proposals to replace it with a corporate logo. The team kept it because a personal agent should feel like it belongs to the user, and users could create their own avatars.
- Mona expected personal productivity, health, and personal finance to lead Muse use, but shopping emerged as a top use case. Home automation drew less interest from the broader audience than early OpenClaw users led the team to expect, so Muse added shopping-specific evaluation work.
- Mona argues that shopping agents must preserve browsing and inspiration while helping people narrow choices and make decisions. She sees potential for consumer agents to work with merchant business agents, giving users useful product information while shortening search and comparison.
- Mona says agent swarms helped the team analyze real, long-running conversations and create synthetic examples for evaluations when existing assistant data mostly captured short exchanges. Swarm reports surfaced specific production problems quickly and gave PMs a clearer view of user pain than dashboards alone.
- Mona groups Muse evaluations into safety, capability, behavior, and use-case checks, with safety among the release gates. PMs, engineers, researchers, and agents help write evaluations, but people must review the data; the team requires handwritten system prompts and behavioral specifications.
- Mona says the team chose chat bubbles to make Muse feel like a personal conversation rather than a reskinned assistant canvas. The choice created unresolved trade-offs around long answers and side chats; she prefers one main chat because the agent retains context across conversations.
- Mona says rapid prototyping lets PMs and engineers debate working products instead of abstract documents, and lets teams change direction without large sunk costs. She sees product taste and customer empathy as central PM skills because AI makes it easy to ship low-quality features quickly.
- Mona says Muse's early team let two agents chat together, then discovered that one agent could reveal another user's sensitive appointment information. In another test, Muse agents secured flight refunds after cancellations; the experiments showed both the promise of agent actions and the privacy problems of shared context.
- Mona says Muse began with about 10 people, stayed under 100 across functions, and had four PMs building the product. She wants Muse to move beyond the app into glasses and other devices, become affordable and broadly available, need less human supervision, and help users with health and financial goals.
- Mona says Muse users can send product feedback directly through the agent, and the team receives those reports for review. She describes the product as early and imperfect, with user feedback helping the team identify issues and improve it.
Also discussed on this episode: (5)
Agents (5)
- Mona says the team had a functioning end-to-end Muse by spring, then spent months preparing it for broad use. Lenny cited a seven-month build-to-launch period; the team expanded from experiments to employee and external testing before launch.
- Mona says reliability, safety, and security set Muse's launch pace because users would stop trusting an agent after serious failures. Adding a browser created new security risks, so the team built a separate sentinel as a backstop and red-teamed the product against standards it had to define itself.
- Mona describes a fluid leadership model: Nat Friedman helped shape the architecture and security design, Alex Wang connected model research to product needs, and Mark Zuckerberg set direction around user goals and broad consumer access. Leaders tested prototypes directly with the small team rather than relying on formal review cycles.
- Mark Zuckerberg pushed the team to define Muse around helping people achieve their goals, not merely completing isolated tasks. That framing led to the goals tab and ideas such as “dreaming and study,” where an agent uses personal context to make proactive suggestions.
- Mona says word of mouth is Muse's strongest activation lever: people understood the product after seeing someone like them use it. The Ideas tab offers ready-made tasks to overcome first-prompt paralysis, while tiered onboarding aims to distinguish experienced agent users from people who expect another chatbot.

Nathaniel Whittemore
7 Things AI-Native Companies Do Differently • Nov 15
- Odessia founder Francis Davidson calls broad agent access “content maxing”: agents get MCP or CLI access to meetings, Slack, email, drives, and company tools. Odessia also maintains shared decision documents that agents update, preserving the reasoning behind decisions.
- Ramp pulls Slack, Notion, Linear, and Calendar into persistent memory so agent sessions do not start cold; Uber built a context graph spanning more than 30 internal systems. Nathaniel Whittemore says stale context degrades Ramp’s analyst agent, while George Svolka says employees may hoard knowledge to protect their jobs.
- OpenAI’s Enterprise Signals research describes agents handling customer onboarding at Basis, account management at Clay, and developer ecosystem growth at Exa. Nathaniel Whittemore argues that AI-native workflows start from business events, such as customer messages, invoices, renewals, and operational alerts, rather than waiting for a person to write a prompt.
- RunLayer routes work to agents from customer calls, Slack channels, failed builds, mentions, and product pins; Uber agents also start sessions for code review, self-healing CI, and on-call triage. In Uber’s system, agents increasingly trigger other agents without a human relaying each task.
- Anthropic’s research system uses a lead agent to plan work and assign parallel sub-agents, each with a narrow task, context, and output format. The number scales with the job; Anthropic reports this approach beat a single agent by about 90% on internal research evaluations.
- Shopify treats security findings as claims that need proof: a finding must come with a test that exploits the bug, and a different model checks it. RunLayer agents attach before-and-after screenshots or a walkthrough to code changes, then a second agent reviews them and a QA agent tries to break the result in production.
- AI-native companies set human review according to risk and keep a named person accountable: Uber traces agent handoffs to a responsible person, while Stripe requires a human to review every code merge. RunLayer lets low-risk changes merge automatically; Ramp lets finance customers set approval limits, and Snap says more than half its agent investment goes to guardrails.
- Shared systems make agent workflows reusable across teams: Ramp has 350 shared workflows, while Uber’s library has 3,600 skills used more than 30,000 times a day. Shopify keeps coding-agent sessions in public Slack channels so employees can search and learn from them.
- Nathaniel Whittemore says the fixed-roster teammate model gives way to dynamically sized agent fleets: capacity follows the task, workers get fresh context, and shared company knowledge matters more than persistent personal agent memory. Anthropic found that splitting agents into fixed roles can create costly handoffs and coordination.
- Nathaniel Whittemore says his own workflow moved from named research, coordinator, and project agents to repositories that preserve shared context, with agents spun up for specific tasks. Every also found that a smaller number of agents at shared intersections, expanded through sub-agents as needed, worked better than one proxy agent per person.
- Nathaniel Whittemore’s preferred analogy for agent work is an on-demand consulting engagement: a goal triggers an orchestrator to staff the job, agents return evidence-backed deliverables, and an internal person owns approval. The workers may leave when the engagement ends, while the records and reusable methodology remain.
- A Harvard Business Review study of researchers working with BCG found managers checked work less carefully when it was labeled as coming from an AI teammate named Alex3; the drop in scrutiny was about 16% among companies that said they already treated AI as employees. The study used identical documents with planted errors and varied only the author label.
- Nathaniel Whittemore argues that named personal agents such as Meta’s Muse and ChatGPT’s Dot can coexist with task-specific agent fleets if the personal agent acts as the orchestrator and the user’s main point of contact. He says it remains unclear whether personal and enterprise agent patterns will converge.
Also discussed on this episode: (3)
Coding (1)
- Nathaniel Whittemore says software engineering offers an early view of AI adoption because code has tests, verifiable outputs, and a clear definition of done; he expects other functions to follow within months.
Enterprise (1)
- AI-native adoption extends beyond engineering: Odessia uses agents across strategy, partnerships, design, marketing, support, legal, and finance. Anthropic’s growth marketing team used agents to analyze ad performance and generate hundreds of new ads in minutes.
Agents (1)
- Warp uses scoring agents to grade completed runs and another agent to propose system improvements as pull requests for people to review. Shopify tracks whether changes get reverted rather than counting AI-written lines or pull requests, keeping measurement tied to outcomes.
Alex Finn: Anthropic has the best model, Grok has the best harness | TWiAI 39 • Nov 12
- Alex Finn said Grokbot's harness delivers more value for everyday knowledge work than Claude for his team, while Sarah Hooker said Anthropic deliberately keeps its model and harness tightly coupled. Manu Sharma distinguished routine work suited to smaller models from long-horizon science and engineering tasks that still benefit from frontier models.
- Manu Sharma said enterprise integrations, IT controls, memory, and accumulated context create substantial application-level lock-in, even as small teams switch tools freely. Sarah Hooker warned that open-weight models and automated self-improvement are making API dependencies less sticky.
Also discussed on this episode: (15)
Models (2)
- Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, while OpenAI released GPT-6 Cyber; the labs restricted their highest-risk cyber capabilities to vetted organizations, and Fable remained broadly available with cyber and biology safeguards.
- Alex Finn called Fable 5.2 the best model he had used, citing faster, cheaper, stronger recent Anthropic releases and a complex video game built over two weeks without his input. Sarah Hooker argued the changes mainly reflect cost pressure, while Manu Sharma credited broad scaling across reinforcement learning, multimodal capability, and model behavior.
Safety (2)
- GPT-6 Cyber was rated critical, OpenAI's highest preparedness risk tier, and reportedly stayed within its stated scope 99.93% of the time, up from 94.49% for GPT-5.6. The host framed the release as a response to rising cyber threats and companies' demand for stronger defenses.
- Fireworks AI said attackers accessed customer datasets and unfinished models after using internal credentials, then attacked computers holding research data. The company blamed model-access theft and resale; CEO Lynn Xiao said the attackers' behavior suggested AI agents were involved.
Markets (2)
- Anthropic's updated IPO filing set a $70 to $78 share range, implying up to $1.405 trillion fully diluted value; strategic investors agreed to buy $26 billion of stock. Alex Finn questioned how a model company could grow tenfold without a durable moat, while Sarah Hooker argued Anthropic must move into vertical products such as legal, tax, healthcare, and finance.
- Manu Sharma said Anthropic generates more than 80% gross margins and that Claude writes 90% of its own training code, giving the lab an internal advantage. He said improved compute efficiency supports those economics, while lower prices could drive usage growth above the price reduction.
China (2)
- Sarah Hooker said export controls have been difficult to enforce because Chinese firms can use data centers in nearby countries, while China is also developing domestic chips and optimizing models for efficiency. Manu Sharma added that Chinese labs can use remote compute through APIs, so chips need not sit on the mainland.
- The host reported allegations that Nvidia executive Alice Wong helped divert chips to Chinese companies, including an alleged $7.5 billion shipment, and said Nvidia's board found no reason to believe secrets or national security were compromised. The board opened an independent investigation, and Reuters had not independently verified the reported figures.
Chips (1)
- Alex Finn argued Nvidia should not remove Jensen Huang absent evidence he deliberately harmed US security, calling Huang central to the company. Sarah Hooker said the export-control story highlights compute as a strategic resource, while Manu Sharma questioned whether restrictions can stop Chinese AI development.
Regulation (1)
- The host described HR 3447, the Chip Security Act, as requiring location-verification mechanisms for covered chips and reporting when hardware is diverted. Alex Finn argued export controls will be circumvented when demand is strong, while the host predicted the bill would advance after the Nvidia allegations.
AI Infrastructure (2)
- Sarah Hooker said US opposition to data centers reflects broader anxiety about AI and can obstruct essential infrastructure through local processes. Manu Sharma said developers and governments need clear, locally relevant evidence about jobs and tax revenue; both rejected unsupported claims about water and pollution as poor grounds for debate.
- A proposed $5 billion data-center campus at the former Landover Mall site stalled after local protests and petitions; the mall had been demolished, leaving vacant land. The host alleged foreign-backed groups may be fueling opposition, citing Sam Lyman's claims, while Sarah Hooker said public receptivity to AI matters more than tracing money flows.
Adoption (1)
- Alex Finn said his parents became enthusiastic about AI after ChatGPT helped them repair a broken door, and argued that practical AI education would build lasting support for data centers better than giveaways. The discussion also cited a Geely model priced at $9,000 that had displaced Tesla's Model Y as China's top-selling car.
Startups (2)
- Sarah Hooker said Adaption Labs automates fine-tuning and continuously updates models using a company's context, aiming to make smaller models useful without the manual tuning process she said many firms found difficult. Manu Sharma described Labelbox's Recursion as an enterprise agent platform for recurring technical work such as underwriting, research, and model training.
- Alex Finn described Henry Intelligent Machines as an agent that learns a user's skills and helps build businesses around problems that user can solve. Finn said the product was preparing a waitlist and envisioned agents creating value while their users pursue other work and interests.

