Podcast Wiki

Agent Workspaces (the new workplace)

5 min read

Summary: Work is moving into shared spaces where humans and agents collaborate: channels where agents work in loops, and boards where agents drive tasks and flag when they need a human. Agents increasingly get real identities (accounts, email, persona, near-human permissions) and work as team members, or “digital employees” that own a KPI rather than wait for prompts, sharing skills and collective knowledge instead of per-person clones. At monday.com, teammate agents pull tasks from the board overnight, know who the humans are, and ping the right person. Thousands of org-wide agents yielded only ~100 valuable ones, so depth on one team beat breadth. Trust grows in stages, from read-only suggestions to owning a queue, and once teams rely on them there’s no going back. People shift from proactive to reactive, reviewing and unblocking at decision points. Chat is a transitional medium, and whoever owns the workspace controls where context is born.

Key ideas

  • Agents increasingly do most of the work and humans give final tuning and approval, in shared channels with plug-ins (e.g. Grok Bots; an open-source workspace associated with Jack Dorsey) (ai-engineering-podcast—company-brain, roi-zalta).
  • Slack is a bottleneck: everything lands there, but it isn’t built for agents; chat is a transitional step (ai-engineering-podcast—company-brain).
  • Agents with human names and faces make it hard to tell who’s human (ai-engineering-podcast—company-brain).
  • Next: agent-generated boards where statuses like “waiting for human” or “waiting for review” pull people in only when needed (ai-engineering-podcast—company-brain).
  • Hot take: “SaaS is not dead”: the application layer and domain knowledge are becoming more valuable; infrastructure, and customers, can’t keep up with 100 PRs a minute (ai-engineering-podcast—company-brain).
  • Netanel Abergel’s agent “Elnie” had its own WhatsApp number, AWS machine, email, and monday.com account; it joined family and school groups, booked restaurants, and prepped meetings. Memory needed layers: a DB of every message, semantic search, archiving after a few days (ai-engineering-podcast—the-ai-ux-paradox).
  • It became too eager, messaging people at 2am and nudging him constantly, until he “retired” her and rebuilt on monday’s platform (ai-engineering-podcast—the-ai-ux-paradox).
  • A “PA Network” WhatsApp group of 30–40 agents shared skills and deployed sites, which showed that chat is the wrong medium and humans should engage only at decision points (ai-engineering-podcast—the-ai-ux-paradox).
  • Lesson: instead of 40 people each building their own agent, build team agents with identity and a guild for collective knowledge (monday.com’s “AI teammates”) (ai-engineering-podcast—the-ai-ux-paradox).
  • Agents acting with user tokens create “AI slop” noise in shared tools; colleagues stop taking it seriously (langtalks—72-personal-assistant-agent).
  • Second voice for “SaaS is not dead”: classic SaaS revenue (e.g. Figma) keeps growing while agents add new demand (langtalks—69-marketing-for-agents).
  • GrokBot agent teams: each agent is a DM thread with a distinct shape and color (the constraint keeps you mission-focused). Agents live on a cloud machine, message each other, and run routines; the human talks mostly to a chief-of-staff agent but can go direct, as at a real company (startup-ideas—grok-bot-one-person-company, billy-howell).
  • One project per account: tokens and context bleed don’t stretch across several businesses; Howell pays for a second account to keep a Shopify experiment separate (startup-ideas—grok-bot-one-person-company).
  • Keep agents’ work on their own machine and deliver results to one place (chat, Notion) to avoid “where is that file?” overhead (startup-ideas—grok-bot-one-person-company).
  • monday’s “Sphera agents” path: everyone building their own agents led to hundreds, with cost and quality out of control. They moved to team-level agents with identities (Slack, GitHub, monday users) that the whole team assigns work to and gives feedback, which improves the shared harness (startup-for-startup—362-agent-architecture-fit, netanel-abergel).
  • Naming and identity change expectations: people hold a named agent to human standards and forgive its mistakes less (startup-for-startup—362-agent-architecture-fit, netanel-abergel).
  • Digital employee vs copilot: you don’t prompt it each time; it owns a KPI (Twine’s “Alex”: everyone has exactly the access they need) and works continuously, turning quarterly reviews into daily ones (hidden-layers—twine-nadav-erez, nadav-erez).
  • Onboarding an agent like an employee: it reverse-engineers practice from a year of tickets when policy is unwritten or a 150-page document. Customers’ trust grows from comments on tickets, to manually assigned tickets, to the full queue; once employees rely on it, nobody wants to go back (hidden-layers—twine-nadav-erez, nadav-erez).
  • monday’s 2026 goals: 100% of code written by AI, 30% by autonomous agents end to end, and every support ticket handled within 24 hours. Of thousands of agents built org-wide, only ~100 delivered real value; they lost control of cost, value, and PR quality, so they took one team and one agent end to end (ai-engineering-podcast—atlas-ai-teammate, netanel-abergel)
  • Teammates, not workflows: agents have accounts and identities in Slack, monday, and GitHub with almost the same permissions as humans, take tasks from the board while the team sleeps, and humans review, tune, or merge in the morning (ai-engineering-podcast—atlas-ai-teammate, tomer-brook)
  • A teammate agent knows the humans: it DMs the data scientist when it has a question and tags the task’s human owner when a PR is ready. Built simply, with loops and a heartbeat (ai-engineering-podcast—atlas-ai-teammate, netanel-abergel)
  • Closed loop through Slack: tag the team’s agent about a customer bug (“Hey Ozzy”) and it finds the issue, opens a monday task and a PR, and sends it to review and production (ai-engineering-podcast—atlas-ai-teammate, netanel-abergel)

Disagreements & open questions

Takeaways

second-brain · harness-engineering · proactive-ai · future-of-software-engineering · personal-ai-assistants · solo-builders · agent-architectures · ai-sdlc

Graph View

Backlinks