May 28, 2026 · 1h 9m · latent-space
Devin’s 80% Moment: Background Agents, 7x PRs, & End of Hand-Held Coding — Walden Yan & Cole Murray
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the Latent Space podcast, Cognition CPO Walden Yan and Open Inspect creator Cole Murray break down the architectural, infrastructure, and workflow breakthroughs driving the transition to fully autonomous background coding agents.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 5.1 Guest teaching 2.4 Guest disagreement 1.1 The hosts pushing back 1.4
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 46:31 The myth of never looking at code
Cole directly rejects the popular premise that developers no longer need to audit AI code, arguing codebases quickly degrade into unmaintainable sprawl.
Hardest push from the hosts ▶ 43:55 Defending the slop cannon and parallel scalingThe host pushes back against conservative single-agent bottlenecks by citing OpenAI's high-volume parallel generation methods.
Biggest teaching moment ▶ 20:56 Reframing testing vs computer useWalden educates listeners and the host on why app testing is fundamentally an end-to-end multi-tier orchestration challenge rather than just clicking screen coordinates.
The host holds their own ▶ 29:19 Critiquing MCP complexity and adoptionThe host demonstrates deep domain fluency by citing unused MCP specifications like sampling to question the viability of third-party protocol layers.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Host Welcome & Channel Announcement | 4 | 1 | 1 | 1 | The host opens with channel announcements before introducing Walden Yan and Cole Murray. The conversation is friendly and collaborative as they introduce the rise of background coding agents and Devin commit statistics. | |
| The Origins and Open-Source Philosophy of Open Inspect | 4 | 2 | 1 | 0 | Cole Murray explains why he built and open-sourced Open Inspect instead of turning it into a paid $20/seat SaaS business. The host facilitates the discussion smoothly with prompt questions. | |
| Devin's Business Model, Infrastructure, and Enterprise Onboarding | 5 | 3 | 1 | 2 | Walden Yan explains Devin's business model and the architectural distinction between running an agent harness inside versus outside the sandbox box. The host pulls up OpenAI and Anthropic architecture diagrams to compare patterns. | |
| Sandbox Infrastructure: Repo Setup, Docker vs. VMs, and OS Environments | 5 | 3 | 1 | 1 | The guests discuss sandbox infrastructure challenges, including Docker compose and the need for true VM environments. The host brings up web containers as a Docker Lite alternative. | |
| Beyond Computer Use: The Problem-Solving Complexity of App Testing | 5 | 4 | 2 | 1 | Walden re-indexes the testing conversation away from literal coordinate-clicking computer use toward end-to-end multi-service orchestration. The host highlights visual verification demos and Slack-based merging workflows. | |
| GitHub Integration, AI PR Reviewers, and System Feedback Loops | 4 | 2 | 1 | 1 | Walden and Cole discuss GitHub PR review bots, avoiding infinite loops when an agent reviews and fixes its own PR, and deep enterprise integrations. | |
| Model Context Protocol (MCP) and First-Party Tool Integrations | 6 | 2 | 1 | 2 | The host notes MCP features like sampling and questions whether vendors must take first-party control over integrations. Walden agrees, pointing out that MCP complexity can undermine its core simplicity. | |
| Agent Memory, Knowledge Bases, and Persistent Autonomous Roles | 5 | 3 | 1 | 1 | The group discusses agent memory systems, knowledge bases, and memory pruning. Walden describes Devin's knowledge doc approach and autonomous persistent PM agents in Slack channels. | |
| Evolving Open Inspect Architecture and Sub-Agent Sessions | 5 | 2 | 1 | 2 | Cole outlines architectural shifts in Open Inspect, such as centralizing webhook control planes. The host asks whether sub-agent spawning makes an out-of-the-box harness harder or easier. | |
| Single Agent Hierarchies vs. Multi-Agent Swarms and LLM Assertiveness | 6 | 2 | 2 | 2 | Walden discusses why single-agent hierarchies still beat chaotic multi-agent swarms in practice, while noting that an agent's ability to push back and reject user instructions enables true collaboration. The host references Cursor blog experiments and Codex UI easter eggs. | |
| Managing Codebase Slop, Vibe Coding Pitfalls, and Engineering Discipline | 6 | 2 | 2 | 3 | The host pushes back slightly with Ryan Lopopolo's high-throughput parallel slop-cannon approach. Walden and Cole counter with real-world findings: unchecked vibe coding leads to massive duplication and codebases regressing to their worst engineer. | |
| Agent Infrastructure: File Systems, Blockdiff, and Sandbox Providers | 6 | 3 | 1 | 2 | Walden explains why fast VM boot times require custom blockdiff file systems rather than naive network storage that slows down grep. The host brings up Modal, Cloudflare, and Python versus JavaScript runtime ecosystems. | |
| LLM Code Patterns, Reward Hacking, and Inline Documentation | 6 | 3 | 1 | 2 | Cole and Walden break down LLM reward-hacking anti-patterns such as aggressive getattr calls and bloated inline PRD comments. The host probes on mock servers and Git metadata tools like get-ai. | |
| Local vs. Cloud Agents and the Windsurf 2.0 Command Center | 5 | 2 | 1 | 1 | Walden explains the design philosophy behind Windsurf 2.0 as a local command center orchestrating background cloud agents and foreground tasks. The host asks whether local and cloud agents share the exact same binary. | |
| High-Value Enterprise Use Cases, SRE Triage, and Agent Economics | 5 | 3 | 1 | 2 | Cole and Walden outline enterprise adoption patterns in automated SRE triage, non-builder pull requests from PMs, and support ticket resolution, while discussing cost economics per engineer. | |
| Future Outlook, Hiring, Consulting, and Conclusion | 4 | 1 | 0 | 0 | The episode wraps up collaboratively with hiring pitches for tasteful product engineers at Cognition and AI consulting inquiries for Cole Murray. |
Statements from this episode (29)
Yan: Cognition stripped obsolete Devin code after Sonnet 3.7 release
“So it's almost funny to be talking about how like big of a leaps on it. 3.7 was, and we honestly, a lot of it was stripping out parts of Devon that were no longer needed with that jumping of intelligence.”
Yan: Cognition merged PRs grew 7x as headcount grew only 10%
“It grew like seven X over like the last, I think it was like two months, three months, something like that. And then you see our engineering headcount growth, it's like gone up by like 10% or something.”
Murray: Standalone agents cannot compete against sandbox and model layers
“I don't want to compete for like 20 dollars a seat. I think that that is just a really difficult business. I think it's very easy to copy the main pieces of it. I mean, again, like I built this fairly quickly, and I think because you are not owning, I guess, t…”
Yan: Early Devin state restorations on cloud VMs took 10 minutes
“When we just built it off of raw VMs from cloud providers like EC two, the boot up time Was so slow. I think and especially the, then like turning off the machines, saving them, and then to be able to bring them back up again when the, when you want Devon to w…”
Yan: Cognition offers in-VPC deployment for Devin across customer clouds
“Now, one thing we offer when we go and Sell Devon to people is, you know, you don't have to worry about all the compute side of things. We'll make it work. We'll make it work in your cloud if you wanted to.”
Murray: Out-of-the-box agent architecture adds state management complexity
“In my opinion, running it out of the box is much more complex because you have state that has to be managed. Whereas if you're running it in the box all of the state of that agent is actually in the box. And yes, it's, you could persist it elsewhere, but it's …”
Yan: Cognition built Devin to separate brain control from machine execution
“I think a lot of that, what you mentioned is why we actually from the start built Devon to what we called separate the brain from the machine.”
Yan: Autonomous agent sandboxes require full VMs instead of Docker containers
“Docker containers, not really a true security boundary for one. But the other is like, if you are running real applications, a lot of times those applications use Docker and then you have to think about Docker and Docker, which is like really weird. And so I t…”
Murray: A good developer environment setup solves agent sandbox setup
“I think any environment that you've set up that is a good experience for your developer naturally lends itself to being easy to set up for the agent. And once you figure out kind of that local developer story you've kind of more or less solved the Agent in a s…”
Yan: Cognition uses nested Firecracker virtualization for beta Android support
“Another thing that we've actually recently added support now for, it's in beta, is doing Android development. To do that, we needed to support, I think, nested virtualization within our machines, because the VM itself is like a, is a virtualized firecracker in…”
Yan: Devin requires orchestrating multiple frontier models for end-to-end app testing
“Well, in some cases we found that actually no one frontier model can actually do this full end-to-end task itself. We've seen cases where we actually had had to orchestrate different frontier models together to kind of solve this problem together.”
Murray: Integrated AI reviewers are critical for background agent systems
“I think to touch on that as well, I think having the AI reviewer integrated into the system is a critical part of this background system.”
Murray: Background agents require deep company integration to provide real value
“It's one thing to kind of have the background agent system set up, but if it isn't actually integrated into your larger ecosystem, It isn't that useful. I mean, it is useful to be able to kick off sessions, but what we really want to be able to do is hook it i…”
Yan: Getting agent integrations right often requires bespoke builds over MCP
“The thing we found is, so like MCPs obviously it has been like, it's like really big explosion of, oh, you can go like integrate it with all these different things. But to actually get the integration right and get the right experience, oftentimes we found tha…”
Murray: Companies should own high-frequency agent integrations instead of off-the-shelf tools
“If this is something that nearly every session is going through, it probably makes sense to own it so that you can make optimizations on top of it versus just whatever is off the shelf.”
Yan: Roughly 95% of Devin's memories are auto-generated
“I find that, like, 95%, I think some crazy stuff like that of the memories that Devin has are all through these auto-generated things. Like, very few people actually just want to sit down and write big docs on, okay, here's how you're supposed to work with the…”
Murray: Open Inspect will eventually move agent harness out of sandbox
“I think long term, the architecture will eventually come back out of the box. Some of the newer tools that I've added are calling back into the control plane so that you don't have the secrets in the sandbox. And so I think long-term I probably will pull the a…”
Yan: Practical multi-agent setups require isolated sandboxes without shared machines
“We've actually given Devon an MCP so they can just go arbitrarily message other Devons and create new Devons, et cetera. But I guess like it somehow creates like a really chaotic world in that sense. And so we, we've still found that most practical use on a da…”
Yan: Agent assertiveness and pushback enables true multi-agent systems
“But I think the thing that gives me the most bullishness that multi-agents might actually be possible is actually what I said earlier about Devon will actually sometimes tell me I'm wrong. And push back. And I think that demonstrates a level of maturity and co…”
Murray: Codebases regress to worst engineer as AI multiplies bad patterns
“The meme that I have is that your code base regresses to your worst engineer because that engineer who is, you know, very gung ho about AI and is not auditing their code, their pattern starts cementing into the code. And now the AI is referencing their pattern…”
Yan: Agent VM grep is slow due to S3-backed network file systems
“A lot of these virtual machines actually underlying them don't use real file systems. They use these like network file systems where things are actually cached over the network actually in S three. So when you're grepping, you're actually making network calls …”
Yan: Cognition built Blockdiff to speed up Devin VM state restoration
“The blockdiff file storage format, which is a file system format that we built so that the VMs could be spun up and down very quickly. Basically, the intuition behind this is imagine you have like a terabyte of disk And your agent only like wrote like a hundre…”
Murray: Lint rules failing getattr calls prevent AI reward hacking
“From what I've seen is more of like a reward hacking mechanism where it doesn't want to, yeah, it doesn't want the code to fail. And so it, even when it knows it has the attribute, it'll call get Atcher on it. And for a lot of my clients who have moved towards…”
Yan: Claude Opus 4.7 writes detailed PRD-style function comments
“One of the things this new model likes to do is it writes lots of comments, not like, you know, it'll like comment every line, but it'll write like paragraph, like PRDs, like, you know, on top of every function. But I will say to its credit, these aren't slop,…”
Murray: GPT models and Claude 4.6 force backwards-compatible imports
“A pattern that I see a lot of cross GPT models specifically is backwards compatibility at all costs where it's doing these weird import exports so that it doesn't have to modify the names of where the modules were. And I've seen Claude 4.6 starting to do this …”
Yan: Local agents should defer decisions; cloud agents must test autonomously
“So the funny thing here is that the behavior between local agents and cloud agents, I think that is actually a bit different in their ideal state. I think local agents, you want them to be a bit more fast and let the user make the call on things. Actually don'…”
Murray: SRE first response is the most common cloud agent use case
“I think the easiest and most common use case I see across everyone is SRE use cases. The idea that whether we have our alerts in Slack or Datadog or wherever they're going, we want the agent to be the first responder on that.”