The Future Live: Agents at Work, Persistent Memory and AI for Health

Matthew Berman and Nick Wentz open the March 13, 2026 show by discussing the web infrastructure that agents need, then speak with leaders from Slack, Giga, Letta and Microsoft AI.

Wentz describes a thesis that agents will increasingly use software on people’s behalf, while pricing, web standards and enterprise systems were built for human users. In this view, structured data, current documentation and APIs become important product features.

Berman says many business products already expose APIs, whereas consumer websites have spent years blocking bots. Browser control can be slow and unreliable, making tasks like shopping, paying a bill or reading a subscribed publication difficult to delegate.

Advertising creates a business incentive to resist agent access, Berman argues: when an agent visits a page, a human may not see its ads. The hosts expect better browser control and negotiated interfaces to develop in parallel, though the timing remains uncertain.

Wentz asks how agents will cope with undocumented procedures. Berman argues that documenting workflows and organizational knowledge already matters when employees leave, and gives agents context that otherwise has to be rediscovered.

The hosts compare OpenClaw with Perplexity’s Computer and Personal Computer. Wentz sees a path to broader accessibility; Berman thinks the setup and trial and error could still intimidate nontechnical users, while work use may be easier to justify.

Slack executive Rob Seaman describes the company’s ambition to provide an operating system for an agentic enterprise: a shared place where people and agents communicate, work with existing apps and access organizational context.

Asked about channel overload, Seaman says AI can summarize and sort information, and Slackbot can find a conversation by its topic and participants without the user remembering the exact channel name.

Berman presses on privacy. Seaman says the Slackbot discussed in the interview is personal: it accesses only Slack content the user can access and connected services under that user’s authorization. The user chooses whether to share its results.

Seaman distinguishes that personal bot from other agents already participating together in channels. Slack was prototyping a shared Slackbot experience, he says, but had no plan at that point to make the personal bot multiplayer.

Seaman reports that Slackbot reached 50,000 weekly active users inside Salesforce in its first few weeks, helped by coworkers sharing useful prompts. He describes a skills library being developed so individuals and teams can reuse and share procedures.

Slack has to preserve the core messaging experience while accommodating agents, Seaman says. If agents overwhelm the conversations people value, the platform loses the human activity that makes it useful in the first place.

For a practical example, Seaman asks Slackbot to group feedback from a feature channel by sentiment and retain roles while removing names. He can then turn positive feedback into draft posts or negative feedback into a product brief and sprint plan.

He also describes a pricing team combining a proposed price change with customer records and Slack context to suggest account-specific plans. These are examples from his company, not independently verified measurements of impact.

On cost, Seaman says Slack had included Slackbot in plans with usage limits rather than relying solely on consumption pricing. He expects some expensive features could eventually be reserved for higher tiers or priced by usage.

Giga cofounder Varun Vummadi explains why the company moved from enterprise model fine-tuning to customer support. Faster, cheaper base models changed the fine-tuning business, while support was a recurring customer need without a satisfactory solution.

Giga builds support agents for voice, chat and email. Vummadi says customers can bring transcripts and policies, and choose between frontier and fine-tuned models; latency and regulatory preferences can affect that choice.

Asked about a purported 98% resolution rate, Vummadi says deployments often start around 15–20%. Policies may be incomplete or contradictory, so Giga uses strong human-agent transcripts and asks customers to approve suggested changes.

When a call is transferred, the company compares what the person did with what the AI could not do, then proposes policy changes. Missing APIs or permissions can prevent automation, so integrating the tools matters alongside improving the model.

Vummadi says some support workers help manage policies and improve agents, while acknowledging that companies also adopt automation to reduce operating costs. The interview does not establish a universal employment outcome.

He argues that forward-deployed engineers can bottleneck enterprise deployments when they manually translate transcripts into policies and repeatedly test changes. His company’s bet is to automate much of that reactive implementation work.

Berman pushes back that engineers also discover poorly articulated customer problems and educate teams. Vummadi agrees complex, novel problems still need that human work, but thinks well-defined support automation can be productized more extensively.

Vummadi offers examples of outreach to inactive customers and an agent contacting a delivery recipient while helping a driver. He says AI can make some proactive support practical, but the examples are company claims rather than controlled evidence.

On customer trust, he argues that callers should retain a clear path to a human. Demonstrating that an agent can solve a specific problem is more useful than trapping people in an automated loop.

Vummadi considers OpenClaw costly and worries about security and sustained use. Berman says he still uses it heavily and sees rough edges as well as inspiration for competing products; both are expressing personal observations.

Letta cofounder Charles Packer says his work, from MemGPT to Letta, addresses how agents retain and use knowledge beyond a model’s fixed training weights. He compares memory management to an operating system or a file system.

More capable computer-use models allow agents to edit their own memory, Packer argues. Yet memory stored only on one device raises durability and portability questions: what happens if that machine fails or the user changes devices?

Packer explains sleep-time compute as work an agent performs between interactions. Simply triggering frequent background activity can produce irrelevant text or loops; the challenge is finding useful work rather than keeping the model busy.

He says Letta Code launches background reflection agents periodically to review recent activity and write useful memories. A good reflection should extract durable insights rather than merely record mundane events already available in logs.

Packer separates short-term context limits from long-term personalization. Larger context windows may ease compaction during a task, but he doubts a new model release alone will give an assistant years of user-specific experience.

Rather than treating memory as a simple plug-in, he proposes agents that curate it and cooperate with specialized coding agents. He sketches both a memory agent attached to another tool and a personalized primary agent that delegates coding work.

The hosts ask whether accumulated memory creates a competitive moat. Packer predicts it could raise switching costs as model differences narrow, although that is a business forecast, not a demonstrated inevitability.

Packer also warns that accumulated memories may contain highly sensitive information or inferences about people. He raises data ownership, portability and the consequences of leaks as issues for individuals and potentially national security.

Microsoft AI health vice president Dominic King describes his route from surgery through DeepMind and other health work to Microsoft. He says the company saw substantial existing demand for health information in its consumer products.

King says Microsoft saw around 50 million health sessions or queries daily and cites its deidentified usage report: about 40% sought health information, roughly 10% concerned symptoms or test results, and one in seven conversations concerned someone else. These are company-reported figures from the time.

For mental health, King distinguishes everyday support from more serious signs requiring expert help and immediate risks needing urgent support. He sees AI as a complement to clinicians and emphasizes human connection in behavioral health.

Berman describes using AI to better understand a scan before asking his physician questions. King sees a potential to reduce information gaps and help people participate in decisions, while retaining a role for medical professionals.

Wentz asks about missed diagnoses and guardrails. King says triage must balance missing emergencies against sending everyone to emergency care; Microsoft uses clinicians, expert panels, medical policy, evaluations and ongoing monitoring, according to his account.

King describes a hoped-for health companion combining medical knowledge with personal context, including wearable and health-record data. His projections for adoption and capability are ambitions, not evidence that an AI can replace a physician or reliably diagnose every case.

The discussion turns to prevention. King sees potential in identifying patterns in more continuous health data, but says much work remains to turn observations into appropriate, personalized actions.

At the time of the interview, King says Microsoft had announced a separate health experience and opened a waitlist. He describes a planned phased US rollout with feedback and clinical evaluation; this is a historical account of the March 2026 announcement.

The hosts close by returning to the possibility of more proactive health support, while the interview leaves open how well these systems will perform and how equitably they will be made available.