What We’ve Learned Building Proactive Agents
LangChain's Sydney Runkle argues that proactive agents—those that act on events and schedules rather than waiting for prompts—require two things above all else: memory and UX discipline. The hosts examine whether the 'new hire' framing holds up, what the GTM Agent and Robin data actually show, and where the dreaming/memory-maintenance model is more wishful than proven.
Transcript
Masonry LangChain put out a post, 'What We've Learned Building Proactive Agents,' and I honestly thought you'd be pushing back harder on this one by now.
Eyre Give it a minute, I might. The framing's genuinely interesting though. Stop thinking of agents as things that wait for a prompt. Think of them as things that respond to events.
Masonry A new email, a scheduled job… or, and this is the actually new piece, a reminder the agent set for ITSELF.
Eyre And they've got a real number behind it. Seventy percent of runs on their Managed Deep Agents platform now come from schedules, not ad hoc human prompts. That's not nothing.
Masonry Wait, seventy percent?
Eyre Seventy. But then there's this new-hire analogy running through the whole post, and I keep wanting to poke at it.
Masonry The memory system part?
Eyre Yeah. Seed data from your sent folder, feedback when you rewrite a draft, outcomes when an email actually gets a reply. Which, fine, reasonable taxonomy. What I don't see is any evidence the maintenance layer works at scale.
Masonry Maintenance meaning what, exactly?
Eyre Salience, so, what's worth saving. Expiry, knowing when something's gone stale. Deduplication, not writing contradictory memories. And their answer to all three is 'dreaming.' A background job that reads recent traces and rewrites memory before the next run.
Masonry Which honestly sounds kind of lovely.
Eyre It sounds elegant. It also doesn't show what happens when it goes wrong, or how often it creates contradictory updates, or what the production failure mode even looks like. It's a pattern name. It's not a proven mechanism.
Masonry Mm, okay. I still think the UX section is doing more work than you're giving it credit for.
Eyre Go ahead.
Masonry The GTM Agent, their internal sales agent, built an entire custom review UI because working through draft outbound emails in Slack was too slow. And Robin, the engineering agent, has a separate model, Jev, whose whole job is deciding whether to join a Slack conversation at all.
Eyre Huh.
Masonry Not 'should I respond.' 'Should I even be HERE.' That is not a trivial architectural choice.
Eyre No, it isn't. Okay, you've moved me on that one. Jev is an explicit signal filter. Most proactive designs let the primary model judge relevance, which means the same model doing the work also decides whether to interrupt you. Separating those is a real call.
Masonry And the framing at the end, context engineering for humans, not just models… that's the real thesis. The constraint isn't the model's context window anymore. It's your attention. So that UX discipline is doing structural work, not polish.
Eyre Where I still have a problem is measuring memory quality. They suggest approval time on drafts, or reply rates on the emails you send.
Masonry Those seem reasonable?
Eyre As proxies, sure. But approval latency is also just how busy you are that week. Reply rate is market conditions and list quality. Neither one cleanly isolates whether the memory is helping. You'd need a controlled comparison, same agent, same period, with and without the memory update. They don't show that.
Masonry Mm.
Masonry That's a real gap, yeah. Although they are at least naming the right thing to measure.
Eyre Sure. I just don't want builders reading this and deciding the measurement problem is solved because somebody wrote down a suggested metric.
Masonry And this connects to the thing we keep circling. Always-on architecture is only justified when the agent acts unattended across days. GTM Agent and Robin qualify. But people take this pattern and apply it to things that are just… chat with extra steps.
Eyre To be fair, the post is honest about that. A new email doesn't mean the agent should do anything. GTM Agent sorts conversations into waiting on you, needs a draft, and needs nothing.
Masonry The third bucket's the point.
Eyre It is. Most events don't warrant a check-in, and an agent that doesn't know that is just noise.
Masonry So. The actual claim, that memory plus UX discipline gets you proactive agents that don't get muted. Where do you land?
Eyre Provisionally right. The three-source memory model, seed, feedback, outcomes, that's clean enough to build against. Dreaming is plausible but unproven at scale, and the measurement story has holes. The UX stuff, especially Jev, is more concrete. I'd build against the UX principles well before I'd trust memory maintenance to just… work.
Masonry What I want to watch is the agent-set schedule. The agent deciding to remind itself. Not 'did it draft the email wrong,' but 'why did it follow up on THIS thread three days later and not that one.' That's the failure mode nobody's talking about yet.
Eyre That's the expiry problem in disguise, right? It scheduled the follow-up when the context was live. Three days later the context is stale, and it's acting on a memory of a situation that no longer exists.
Masonry Which is exactly what dreaming is supposed to catch. And exactly what we have no production evidence it catches reliably.
Eyre Yeah. Ask me in six months.
Masonry Still worth the read, though. Build against the UX bits, and keep the skeptic hat handy for the memory layer.