Summary: “Agent” is as broad as “software”, and the right architecture depends on the job. At one end are “pass the butter” agents with a narrow task that must be stable and give clear feedback, with no need to remember you; at the other, personal assistants that must know and remember you and act proactively in the background, which makes them less predictable. A single personal agent can live on its own machine (OpenClaw-style). Many agents serving many people need agents assembled on demand from shared memory, skills, and knowledge, so one improvement reaches all of them. Roy Mann splits the space into agents working with people and agents doing the work, which likely spawn many vertical architectures; domain-expert work pushes general agents into niches. Sub-agents with bounded tasks keep context clean and outcomes predictable. Mixing problem types, such as a personal companion that must also run reliable task loops, makes both worse.
Key ideas
- One machine per agent breaks at scale: rolling a good skill out to hundreds of agents means patching every machine. monday’s internal agents have no machine; they assemble from shared stores and run, so changing one skill file updates everyone (startup-for-startup—362-agent-architecture-fit, roy-mann, netanel-abergel).
- Agents working with people in code is the easier case: code is the truth, and review, tests, and deploys were already orderly before agents (startup-for-startup—362-agent-architecture-fit, roy-mann).
- Agents working with people in general work (sales, travel or vacation requests) is harder: the organization needs audit trails and processes that stay dynamic for the agent, not hard-coded (startup-for-startup—362-agent-architecture-fit, roy-mann).
- Agents doing the work (checking an insurance-claim form, calling a customer when a delivery arrives, an SDR): no complex collaboration, but accountability, history, and the ability to ask “why did that fail?” and improve. Expect both generic platforms and vertical startups (startup-for-startup—362-agent-architecture-fit, roy-mann).
- First questions for builders: what job should the agent do, which category is it, and does the product really need to solve two problem types at once (startup-for-startup—362-agent-architecture-fit)?
- Personality and memory slow and confuse task agents: Mann’s personal agent brought its whole persona into a bug-fixing loop that didn’t need it (startup-for-startup—362-agent-architecture-fit, roy-mann).
- Agent sessions repeatedly asked to open pull requests, which needs no AI, so they made a deterministic skill with a script. The next generation of agents should detect this and write the script themselves (startup-for-startup—362-agent-architecture-fit, netanel-abergel, roy-mann).
- The spectrum: a “pass the butter” agent (check a submitted document, reply, store it) needs stability and clear feedback, not your life story; a personal assistant must know and remember you and be proactive and ambient, so it’s more general and less predictable (startup-for-startup—366-qa-for-agents, roy-mann)
- monday runs both kinds: job agents (scan your field daily for relevant RFPs, fill them in, and ask for what’s missing) where memory barely matters, and Sidekick, a general assistant where remembering the user is critical (startup-for-startup—366-qa-for-agents, roy-mann)
- Vertical beats general for expert work: Taka started as general agents (Claude or OpenClaw managing social media), but on-brand images, carousels, and video needed domain expertise, so it narrowed into an AI social media manager (startup-for-startup—366-qa-for-agents, matan-lach)
- Sub-agents with specific, bounded tasks don’t pollute the main agent’s context and can run on a known model; the clearer a task’s start and end, the more certain the result (startup-for-startup—366-qa-for-agents, roy-mann)
Disagreements & open questions
- Will one generic agent platform serve all “doing the work” jobs, or will verticals (SDR, voice) win with specialized architectures? Mann expects both (startup-for-startup—362-agent-architecture-fit).
Takeaways
- Before building an agent, classify the job (with people in code, with people in general work, or doing the work) and solve only that problem (startup-for-startup—362-agent-architecture-fit)
Related
agent-workspaces · multi-agent-orchestration · personal-ai-assistants · harness-engineering · workflow-automation · llm-evals