The five components of building a "company harness"

Masonry LangChain dropped a post this week — Harrison Chase, five components for building what they're calling a company harness. And I keep coming back to the identity section, because they're doing something that I think most teams just… don't.

Eyre The on-behalf-of versus fixed credentials split? Yeah, that one's genuinely interesting. Most teams pick one and call it done.

Masonry Exactly. LangChain is running BOTH inside the same agent depending on the action. OpenSWE opens your PR under your GitHub handle, but when it's doing background code review it switches to its own app identity. GTM Agent does email as you, Salesforce as a service account.

Eyre Which is actually the right call, technically. The alternative is either everyone logs into Salesforce — which nobody's going to do — or the agent has your credentials for everything, which is the nightmare scenario. I mean, this is basically the fake mustache thing we keep coming back to.

Masonry Please stop making my agent pretend to be me with a fake mustache.

Eyre Still the most accurate description of the problem.

Masonry Okay, so the five things: it lives where people already work — Slack for them — dual-mode identity, scoped memory, sandboxed code execution, and then ambient scheduled runs. That last one is the one I don't think gets enough attention.

Eyre Right, and the GTM Agent example is a good one. You ask it to send you a daily brief, and it creates its own cron to do that. The agent is setting up its own recurring job.

Masonry Which flips the whole mental model. Most people build agents they go talk to. LangChain is saying a large percentage of their runs are ambient — no one kicked them off, they just fired.

Eyre Yeah. And that's where the harness-as-product question gets uncomfortable for me, because all five of these things — the channels, the crons, the sandboxes, the memory, the identity — it all runs on Managed Deep Agents. The recipe and the product are the same thing.

Masonry I mean, sure, but that's also just… honest? They built the thing to solve their own problems and they're telling you what the problems were. I don't think the recipe is fake just because they're also selling the infrastructure.

Eyre No, I don't think the recipe is fake. The five primitives are real. We've been watching people converge on these same components for months — this isn't the first thing to land on identity plus scheduling plus sandboxes.

Masonry Right.

Eyre What I'm less sure about is the 'completely changed how those teams work' framing. That's the part I want receipts on — though I guess the GTM Agent post from March is the closest thing they have.

Masonry The self-healing deployment pipeline piece is also real — there's a whole post on that from April, Vishnu Suresh. After every deploy it detects regressions, figures out if the change caused them. That's not demo material, that's production.

Eyre Okay, that's actually good. The memory section is the one where they're most honest — they basically say it's the least-explored area and they have a lot of work left. Two tiers, global and user-scoped, agents decide what to remember at which level. Fine in theory, but context rot is a real problem and they know it.

Masonry There's a whole separate post on that, actually — the context engineering piece from October fourteenth. The short version is that long-running agents accumulate more than the model can hold and accuracy drops. The two-tier memory is their answer to that, but Eyre, I think you're right that it's early.

Eyre Mm-hm. The Slack breakout feature is the thing I didn't expect to care about. The idea that an agent can spin out of a human thread into its own thread so it doesn't pollute the conversation — that's a small UX call that matters a lot in practice.

Masonry It's the kind of thing that only gets invented by teams who actually ran the agent in Slack for six months and watched people get annoyed.

Eyre Exactly. You don't design that in advance.

Masonry So my read on this whole post: the five components are real, the examples are grounded, the memory piece is genuinely immature, and the product pitch is visible but not disqualifying. If you're a team that's been staring at the same five problems and not naming them — this is useful.

Eyre I'd put it slightly differently. The value isn't the framework, it's the identity section. The rest of it most teams will figure out. The on-behalf-of versus fixed credentials per-action split — that's the thing people get wrong and it bites them later.

Masonry Sixty-forty Managed Deep Agents has real external adoption numbers they can point to by end of Q one twenty twenty-seven. Not just LangChain internal — actual other teams shipping on it.

Eyre I'd take that bet. Though I'd want the definition of 'shipping' nailed down before we shake on it.

Masonry Fair. Alright, Eyre — same time next episode, I guess.