Ex-Uber dev explains his Multi-Agent Workflow

Tech/Agent workflow takeaways (from the subtitles)

1) Shift from “single-player” AI to “multiplayer/agentic” work

• AI is moving from one-off tools/agents to multiplayer agent systems that collaborate in team environments.
• Collaboration should feel like editing a shared artifact (e.g., working on a Google Doc link) rather than repeatedly passing files back and forth.
• Key point: most existing software tools were designed for humans, not agents—agents need different affordances, such as:
  • stronger collaboration primitives,
  • shared context,
  • “team-level” integrations.

2) The bottleneck is adoption and workflow redesign—not just model capability

• Technology is often less limiting than organizational adoption—whether a company commits to becoming “a business of the future” or sticks to legacy processes.
• Even if agents can do long-horizon tasks, companies must still restructure how work is coordinated.

3) Why current “agent collaboration” is messy (and what should replace it)

• Today’s collaboration pattern is described as chaotic:
  • teams pass around files (e.g., MD/HTML),
  • each agent/tool runs in its own setup (multiple “cloud code / cowork / CEX”-like environments).
• Although one “best solution right now” is GitHub repos (as described in the segment), the speaker claims it’s not optimal for agent workflows.

4) Importance of “shared context” as the foundation for multiplayer agent experiences

• Shared context is presented as the “bedrock” of multiplayer agent systems.
• Example mechanism: automatic hydration—an agent continuously ingests data from sources like Slack and meetings to keep a team “master memory” and knowledge graph current.
• This enables questions like:
  • “What did we decide?”
  • “What’s the latest shipping date?”
  • “Catch me up on onboarding highlights”
  • without humans manually maintaining wikis.

5) Product direction: agent “team surfaces” (Slack/Google Docs/email) + normal human permissions

• The agent should behave like a normal teammate, including:
  • an email address,
  • being invokable through Slack,
  • operating inside tools like Google Docs.
• Highlighted limitation: many agents still lack basic team-native identity/integration (e.g., email) or don’t plug into collaborative documents.

6) Tooling/UX principle: “web UI is dying” and “zero setup” matters

• Users don’t want to learn new apps or configure many API keys/connectors.
• The “winning” experience should be:
  • non-intrusive,
  • reliable/high uptime,
  • efficient and familiar (e.g., via Slack).
• The speaker contrasts this with harder-to-use automation/orchestration systems (mentioned as categories similar to “N8n/NA10” / Zapier, and “make.com” style complexity).

7) Example build philosophy: let agents manage context + workflows, not humans

• Agent usage patterns mirror how teams already work, such as:
  • recruiting agents that find engineer blog posts and relevant signups,
  • agents drafting emails and preparing meeting notes,
  • agents taking notes and preparing briefs,
  • recurring routines (scheduled recurring agent workflows),
  • using calendars, transcripts, email labeling, and messaging (iMessage/Discord-style).
• In the company context, the guidance is to encourage early experimentation (“let a thousand flowers bloom”), since different engineers will create different agent workflows.

8) “Self-building wiki / self-updating memory” as a feature set

• The described system continuously builds an always-fresh knowledge base from internal communications:
  • updated learnings,
  • a graph of company activity,
  • ingestion across months of Slack history plus ongoing meeting updates.
• This removes the “wiki goes stale” problem.

9) Deep API pitch: single API key unlocks browsing/research/email-style capabilities

One segment is explicitly promotional for Deep API. Claims include that giving an agent a single API key enables:

• world-class scraping for websites/social media,
• deep research (positioned favorably vs Perplexity),
• browser use,
• agents having their own email address,
• sending emails automatically without user auth/config.

Demonstrated use cases:

1. Extract 100 licensed general contractors from Austin, Texas state licensing records; output as a Markdown table (took ~30 minutes).
2. Analyze news/sentiment for 8 stock tickers, in ~4 minutes; also email results back to the user.

Pricing is described as “fully transparent,” with a low subscription tier (e.g., ~$30) supposedly providing thousands of API calls.

10) Company survival framing: “reinvent from scratch” continuously

• Repeated theme: companies must reinvent because adoption and tooling expectations change quickly.
• Warning: if you don’t, you risk slow decline as competitors update faster.

11) AGI risk perspective (high level)

• They don’t emphasize precise timelines; they expect transformation sooner than many people feel comfortable with.
• Framing: “death or house cat” (existential risk vs a benign-but-nasty-in-between scenario).
• Practical stance: focus on controllable near-term adoption of agentic workflows now.

---

Main speakers / sources

• Flo (named “Flo” in the transcript; likely the founder/product voice discussing agent collaboration, shared context, and “Lindy teammate” / “automatic hydration”)
• David (host/interviewer)
• Deep API (promotional/product source for the “single API key” agent capabilities)
• Lindy.ai / “Lindy teammate” (product/source referenced as the multiplayer agentic scaffold)