Concurrent Delivery — Research Log

Aviator MergeQueue

Fri, 16 Oct 2026 00:00:00 GMT

A monorepo merge queue that runs separate parallel queues per independent set of affected targets, each with its own required checks, plus predictive batching, flaky-test quarantine and stack-aware merging.

Splitting the queue by affected paths removes the single serial lane, which is what limits throughput under agent PR volume. The page names no agent-specific features.

Okteto

Fri, 16 Oct 2026 00:00:00 GMT

Starts an isolated, production-like environment on your own infrastructure for each agent task; agents use the same CLI and API as developers, with shared backing services where wanted.

The platform team sets resource limits and sees every agent environment in one place. Per-developer sandboxes don't give that fleet-level view.

“And we also have a concept of like an agent trajectory store as well. That's more focused…” — KP Sawhney, Google DeepMind

Thu, 15 Oct 2026 00:00:00 GMT

“And we also have a concept of like an agent trajectory store as well. That's more focused around the the coding piece where obviously you can have a huge number of of steps going on there. And it can be really important to kind of diagnose at what exact point looping started happening or you know, the model went off the rails.” — KP Sawhney, Google DeepMind

Sawhney describes an internal store that records the steps of long coding-agent runs, so engineers can find where an agent started looping or went off course. It sits alongside a UI for drilling into every request made by agents on Google’s shared agent back end. The captions don’t label speakers; the attribution to Sawhney is inferred from the panel’s flow. Quoted from YouTube’s auto-generated captions.

“on a per language basis, we have a specific like auto review model that has been um…” — KP Sawhney, Google DeepMind

Thu, 15 Oct 2026 00:00:00 GMT

“on a per language basis, we have a specific like auto review model that has been um fine-tuned on all of our style guides and and and all the rest of it um and previous like good examples of of code.” — KP Sawhney, Google DeepMind

To keep up with reviewing agent-written code at Google’s volume, Sawhney describes a per-language automated review model fine-tuned on the company’s style guides and past examples of good code, with product teams adding their own review prompts. The captions don’t label speakers; the attribution to Sawhney is inferred from the panel’s flow. Quoted from YouTube’s auto-generated captions.

“honestly right now it's kind of brute force with the the quota. So we have we have some…” — KP Sawhney, Google DeepMind

Thu, 15 Oct 2026 00:00:00 GMT

“honestly right now it's kind of brute force with the the quota. So we have we have some real power users at DeepMind and ultimately it gets to a point where it's like, okay, you've got to just stop right now.” — KP Sawhney, Google DeepMind

Asked how DeepMind keeps one power user running around a hundred agents from overwhelming a shared system, Sawhney says the current control is a quota cap that stops usage at a threshold. He presents it as a blunt stopgap rather than a finished mechanism. Quoted from YouTube’s auto-generated captions.

“Now, coding was the easy part. This actually was never the bottleneck. Your context is…” — Davis Palmie, Factory

Thu, 15 Oct 2026 00:00:00 GMT

“Now, coding was the easy part. This actually was never the bottleneck. Your context is laid out in natural language. I bet if you look at your real engineering metrics, what's the median time to review a PR?” — Davis Palmie, Factory

Palmie argues that writing code was never the limiting step; the delay sits downstream, in pull request review, reproducing bugs and debugging. He frames the software factory as putting agents into those stages rather than only into code generation. Quoted from YouTube’s auto-generated captions.

“they work in a sequence. So they don't work in a swarm or parallel. The agents work in…” — Tereza Tížková, Factory

Thu, 15 Oct 2026 00:00:00 GMT

“they work in a sequence. So they don't work in a swarm or parallel. The agents work in sequence and everyone accomplishes something and passes it to the next one. uh we actually found that if you do this you end up with more fresh context and kind of fresh head.” — Tereza Tížková, Factory

Factory’s “missions” are long-running sessions, some lasting weeks, where an orchestrator agent hands work to worker agents. Workers run one after another rather than as a parallel swarm, and each one can still use parallel sub-agents for small tasks such as research. Quoted from YouTube’s auto-generated captions.

“real mission from our customers that run 16 hours and just just to see how important is…” — Tereza Tížková, Factory

Thu, 15 Oct 2026 00:00:00 GMT

“real mission from our customers that run 16 hours and just just to see how important is the validation as well. it takes even 40% of the whole process. … the validators judge code that they didn't write. Uh there's something from the orchestrator agent called validation contract uh that's written before any code is done.” — Tereza Tížková, Factory

In a Factory mission, separate validator agents check the workers’ output against a contract the orchestrator wrote up front. One validator does linting, types and tests, and another clicks through the running app on a virtual machine. In the customer run she showed, validation took 40% of a 16-hour mission. Quoted from YouTube’s auto-generated captions.

“And as more correct diagnoses roll in and in from the agent, your team can start to build…” — Davis Palmie, Factory

Thu, 15 Oct 2026 00:00:00 GMT

“And as more correct diagnoses roll in and in from the agent, your team can start to build trust and an understanding of this pillar. So each step in the software factory needs careful ownership and monitoring before you even think about letting the the pipeline loop.” — Davis Palmie, Factory

Palmie recommends starting with one narrow, checkable task, such as incident triage that engineers review. Each stage gets a named owner and monitoring, and stages are chained into a closed loop only once the agent’s output has earned trust. Quoted from YouTube’s auto-generated captions.

“It's actually our best engineer, the one that hold all their [context] in their brain. ……” — Eyal Blum, Figma

Thu, 15 Oct 2026 00:00:00 GMT

“It's actually our best engineer, the one that hold all their [context] in their brain. … they know where all the pitfalls are. They are like holding together with with like their mental duct tape all the places that agents are not working well … and they get so much burden and and become bottlenecks” — Eyal Blum, Figma

Figma found its senior engineers holding unwritten institutional knowledge that agents lack, so they end up catching agent mistakes and adopting AI the slowest. Figma’s response is to put these skeptics in charge of the roadmap for verification and tooling. Quoted from YouTube’s auto-generated captions.

“Make sure that the plan can be broken down into small parts that can each be verified…” — Eyal Blum, Figma

Thu, 15 Oct 2026 00:00:00 GMT

“Make sure that the plan can be broken down into small parts that can each be verified independently. And my personal way of knowing what is a good size would I want to review that the PR that will correspond to that part?” — Eyal Blum, Figma

Figma engineers may spend a week on a detailed plan, aligned with other teams, before handing it to agents. Each phase has its own validation gate so later phases don’t build on unchecked assumptions. One plan came back as about 20 PRs of 10 to 100 lines each. Quoted from YouTube’s auto-generated captions.

“what we see on our platform at least 13% of PRs had a recorded CI failure. … Which means…” — Gautam Korlam, SonarSource

Thu, 15 Oct 2026 00:00:00 GMT

“what we see on our platform at least 13% of PRs had a recorded CI failure. … Which means in all these cases there is a separate round trip to go either back to the developer machine or for an agent to understand the logs root cause of failure fix it push another commit” — Gautam Korlam, SonarSource

Korlam’s code review platform Gitar (now part of SonarSource) sees PR volume growing about 25% month over month. His mitigations: make CI fast enough to absorb repair round trips (their incremental CI went from about 10 minutes to under one), and shift expensive checks left. Quoted from YouTube’s auto-generated captions.

“So if it's a dependency bump and doc changes, we should be approving them. Any changes…” — Gautam Korlam, SonarSource

Thu, 15 Oct 2026 00:00:00 GMT

“So if it's a dependency bump and doc changes, we should be approving them. Any changes under sensitive stuff like payments or [auth] might need a human approver on top. And then of course you don't want to have the agent who made the change approve its change because that would not be uh zero trust.” — Gautam Korlam, SonarSource

To keep human review from becoming the queue, teams on Gitar write plain-English approval policies: low-risk changes are auto-approved and merged, sensitive paths escalate to a person. Each approval and override is recorded on the PR for audit, since CI logs expire. Quoted from YouTube’s auto-generated captions.

“At that point, I'm not running agents anymore. I've become the scheduler, deciding who…” — Kyle Jaejun Lee, KRAFTON

Thu, 15 Oct 2026 00:00:00 GMT

“At that point, I'm not running agents anymore. I've become the scheduler, deciding who does what. I'm the memory, holding what every one of them is doing, and I'm the reviewer, checking all of it. One human, three roles, six contexts. It does not scale.” — Kyle Jaejun Lee, KRAFTON

With about six agents running at once, Lee found himself acting as their scheduler, shared memory and reviewer, and his attention became the limit. He responded by arranging the agents in a hierarchy where each sees only a scoped slice of context, so he reviews only what reaches the top. Quoted from YouTube’s auto-generated captions.

“When two machines point at the same directory and both change it, you get conflicts. The…” — Kyle Jaejun Lee, KRAFTON

Thu, 15 Oct 2026 00:00:00 GMT

“When two machines point at the same directory and both change it, you get conflicts. The same context quietly diverged, so I separated it. Per-machine directories for machine-specific state, and the shared stuff only changes through a pull request.” — Kyle Jaejun Lee, KRAFTON

When agents on two machines edited the same context directory, their state quietly diverged. Lee split it into per-machine directories and lets shared context change only through pull requests. Quoted from YouTube’s auto-generated captions.

“if you have exhaustive verification then you can ship code at the speed that your agents…” — Gabriel Spencer-Harper, Meticulous

Thu, 15 Oct 2026 00:00:00 GMT

“if you have exhaustive verification then you can ship code at the speed that your agents write it. If you don't, then someone somewhere at your organization is spending time verifying code and that is now your new bottleneck” — Gabriel Spencer-Harper, Meticulous

Meticulous, which sells this, records real user flows in non-production environments and replays them against each PR with mocked network calls and a deterministic browser. Isolating each replay this way removes race conditions, so the replays can run in parallel with few flaky results. Quoted from YouTube’s auto-generated captions.

“imagine you have a while loop and just forking processes in one tool call, right? So now…” — Abhishek Bhardwaj, OpenAI

Thu, 15 Oct 2026 00:00:00 GMT

“imagine you have a while loop and just forking processes in one tool call, right? So now you've kind of become a bad neighbor or a noisy neighbor and you've kind of brought down the node and no other tool calls can run, right?” — Abhishek Bhardwaj, OpenAI

Bhardwaj works on the sandbox infrastructure that runs untrusted code for ChatGPT and Codex web. Plain fork-and-exec lets one tool call starve a shared node. Containers add cgroups for resource limits, but processes in them still share the host kernel with other users’ sandboxes. Quoted from YouTube’s auto-generated captions.

“Like in the end, everyone always wants a VM because they tried everything. They tried…” — Abhishek Bhardwaj, OpenAI

Thu, 15 Oct 2026 00:00:00 GMT

“Like in the end, everyone always wants a VM because they tried everything. They tried containers, G visor, V8s. … Just please use micro VMs from the start.” — Abhishek Bhardwaj, OpenAI

His recommendation for agent sandboxes at fleet scale is hardware-virtualized microVMs (Firecracker, Cloud Hypervisor), with incremental disk snapshots so a sandbox can be restored on another node when hardware fails. He accepts the performance cost over the risk of a breach between tenants. Quoted from YouTube’s auto-generated captions.

“They have tens of thousands of developers, thousands of thousands of repositories. … they…” — Robert Brennan, OpenHands

Thu, 15 Oct 2026 00:00:00 GMT

“They have tens of thousands of developers, thousands of thousands of repositories. … they have seen a 30x improvement on time resolutions for these [CVEs] by doing uh orchestration at scale.” — Robert Brennan, OpenHands

For one OpenHands customer, every new vulnerability starts an agent per repository that finds affected code, updates the dependency, fixes breaking API changes and opens a PR. The owning team only has to validate and merge it. Quoted from YouTube’s auto-generated captions.

“A slightly more sophisticated thing would be to create a dependency tree. … you start with…” — Robert Brennan, OpenHands

Thu, 15 Oct 2026 00:00:00 GMT

“A slightly more sophisticated thing would be to create a dependency tree. … you start with like the leaf nodes in your dependency graph right you start with maybe your utility files get those migrated over um and then anything that depends on those you know it's going to have those those initial fixes in place” — Robert Brennan, OpenHands

OpenHands splits large refactors into PR-sized batches that a single agent can finish in one pass. Batches run in parallel in dependency order, and a batch waits until the batches it depends on are verified and merged into a shared integration branch. Quoted from YouTube’s auto-generated captions.

“you can also have the agents basically share context with each other through a file like…” — Robert Brennan, OpenHands

Thu, 15 Oct 2026 00:00:00 GMT

“you can also have the agents basically share context with each other through a file like agent MD. … maybe they send a pull request into the file as they learn new things. Uh downside here is that sometimes agents will try and learn unimportant things. … doing some kind of human review seems to help.” — Robert Brennan, OpenHands

When ten parallel agents all hit the same problem, OpenHands wants the fix shared between them. Brennan compares sharing all context, having a human paste in hints, letting agents edit a shared AGENTS.md with human review, and agent-to-agent messaging, which adds non-determinism. Quoted from YouTube’s auto-generated captions.

“But it is quite different to say, I'm going to now also give my agents the ability to…” — Ben Dicken, PlanetScale

Thu, 15 Oct 2026 00:00:00 GMT

“But it is quite different to say, I'm going to now also give my agents the ability to deploy changes to the database or make changes to how things are configured there or to shard my database, right? because that's something if it gets wrong, it's generally not as simple as clicking a button to roll it back, right?” — Ben Dicken, PlanetScale

Dicken separates agents writing application code from agents deploying schema or configuration changes to a production database. A bad database change is hard to undo and can take the whole application down, so it needs its own safety mechanisms. Quoted from YouTube’s auto-generated captions.

“I think there was a a tweet a couple of months ago from GitHub that basically said like…” — Ben Dicken, PlanetScale

Thu, 15 Oct 2026 00:00:00 GMT

“I think there was a a tweet a couple of months ago from GitHub that basically said like their traffic had like two or three or 4x from what already was an extremely high load. And pretty much all of that was attributed to AI agents, right?” — Ben Dicken, PlanetScale

Dicken recalls, secondhand, a GitHub post about traffic multiplying over an already high load because of AI agents, as an example of agents overloading shared infrastructure. His mitigation is to tag traffic by source and give each segment a resource budget, so a surge from one source degrades gracefully instead of taking the system down. Quoted from YouTube’s auto-generated captions.

“we also have the APIs and the CLI to be able to let agents automate this. So when they are…” — Ben Dicken, PlanetScale

Thu, 15 Oct 2026 00:00:00 GMT

“we also have the APIs and the CLI to be able to let agents automate this. So when they are doing things and automating code and branching and making pull requests and doing code review, there can actually be a mirror database state that is working in sync with all of those things, right?” — Ben Dicken, PlanetScale

Dicken describes giving agents API and CLI access to database branching, so each agent branch or pull request gets a matching database state instead of every agent sharing one schema. Changes are merged through a deploy request and can be reverted. Quoted from YouTube’s auto-generated captions.

“biggest limitation factor that we see with this is agent memory is still very limited to…” — Kamalakannan Nandagopal, Postman

Thu, 15 Oct 2026 00:00:00 GMT

“biggest limitation factor that we see with this is agent memory is still very limited to an individual user, and no standard patterns have now emerged yet in terms of sharing agent memories across your broader teams or across your entire engineering organization.” — Kamalakannan Nandagopal, Postman

Postman runs more than 150 microservices with thousands of REST endpoints across hundreds of repos. Instead of per-user agent memory, it built a shared API context graph, grounded in code and production traces, that maps every service, endpoint and consumer for its coding agents. Quoted from YouTube’s auto-generated captions.

“there were new consumers between the time where the eval was written and the time it the…” — Kamalakannan Nandagopal, Postman

Thu, 15 Oct 2026 00:00:00 GMT

“there were new consumers between the time where the eval was written and the time it the evaluation was run. So the context graph is not only important to populate, but it's equally important that you keep it up to date, and any delays in syncing your context graph with the ground truth uh creates gap in the result.” — Kamalakannan Nandagopal, Postman

An impact-analysis eval failed because new API consumers had appeared since the graph was built. Postman found about 75% of PRs in its top 10 repos touch an API surface, so stale shared context quickly leads to wrong answers and lost developer trust. Quoted from YouTube’s auto-generated captions.

“being in this isolated environment where I'm working with an agent, maybe I start with my…” — Matt Dailey, Ref

Thu, 15 Oct 2026 00:00:00 GMT

“being in this isolated environment where I'm working with an agent, maybe I start with my vague idea and I'm exploring it and asking questions, but decisions are being made in that like that chat that are are not shared with my team that are going to disappear” — Matt Dailey, Ref

Ref argues chat sessions are private and short-lived by design. It suggests keeping state in a shared doc that every agent starts from, so agents stay stateless, the team can review the decisions, and there is a durable decision log. Quoted from YouTube’s auto-generated captions.

“too many PRs to merge. This is the like classic first problem you hit when you start…” — Matt Dailey, Ref

Thu, 15 Oct 2026 00:00:00 GMT

“too many PRs to merge. This is the like classic first problem you hit when you start adopting AI as an engineer. … and this is obviously magnified when the whole team does this. Merge conflicts, merge queue breaks down, things get bad.” — Matt Dailey, Ref

Dailey calls the team-level effect of sudden AI speed “velocity sickness”: more PRs than anyone can merge, people moving in different directions, and repeated work. His fix is to align on key decisions in shared docs before agents implement, which moves the review point earlier. Quoted from YouTube’s auto-generated captions.

“There are different coding standards being applied in different places by different agents…” — Dan Adler, Sourcegraph

Thu, 15 Oct 2026 00:00:00 GMT

“There are different coding standards being applied in different places by different agents across the codebase. There's duplicated code proliferating. You you already have a library that does this, but your agent doesn't know that.” — Dan Adler, Sourcegraph

Adler’s point is about enterprises with tens of thousands of repositories, far too much code for an agent to fit in context or grep in real time. Sourcegraph’s answer is a code graph and search layer that agents can query across every repo, plus batch changes that roll a fix out to thousands of repos at once. Quoted from YouTube’s auto-generated captions.

“when you say, "Hey, you know, wipe this agent database." and it finds a token on your…” — Arjun Singh, Superconductor

Thu, 15 Oct 2026 00:00:00 GMT

“when you say, "Hey, you know, wipe this agent database." and it finds a token on your laptop that it can use and it thinks it's working with staging, but actually it's production and now it just deleted everything.” — Arjun Singh, Superconductor

Superconductor moved all agent work off laptops into isolated cloud environments. Each one gets only the credentials it needs and a configurable network allowlist, which also lets non-technical teammates trigger real work safely. Quoted from YouTube’s auto-generated captions.

“essentially 100% like 99.9% of our pull requests are like heavily agent generated. We know…” — Arjun Singh, Superconductor

Thu, 15 Oct 2026 00:00:00 GMT

“essentially 100% like 99.9% of our pull requests are like heavily agent generated. We know that quality and reliability and security are really important, so we still have humans look at everything. We have agents help with it all, but everything's human reviewed.” — Arjun Singh, Superconductor

Superconductor’s small team used about 1.5 billion tokens in a month and benchmarks agent harnesses on its own past PRs to pick a default. Even with nearly all code agent-written, a human still reviews every PR. Quoted from YouTube’s auto-generated captions.

“if you just kept everything serialized and merged them one at a time so as not to break…” — Gautam Korlam, SonarSource

Thu, 15 Oct 2026 00:00:00 GMT

“if you just kept everything serialized and merged them one at a time so as not to break your uh main branch you would take a really long time up to days so 20 days in this case to make sure that everything was green.” — Gautam Korlam, SonarSource

At Uber, thousands of daily changes to one monorepo with long builds made one-at-a-time merging impossible, while letting everything merge kept main red. Uber’s SubmitQueue tested likely combinations of pending changes in parallel and kicked out the ones predicted to break trunk. Quoted from YouTube’s auto-generated captions.

“U-Review does like around 25,000 comments a week. And uh we get 10% of them actually get…” — Ameya Ketkar, Uber

Thu, 15 Oct 2026 00:00:00 GMT

“U-Review does like around 25,000 comments a week. And uh we get 10% of them actually get some feedback. And only 4% of the PRs actually get some negative feedback. Uh we also saw that um the overall addressal rate was uh around 67%” — Ameya Ketkar, Uber

uReview routes each change to generators by risk and complexity, then rates, filters and deduplicates comments before posting. Uber tracks whether developers act on each comment (the “addressal rate”) and reports costs down 60% against a naive setup. Quoted from YouTube’s auto-generated captions.

“Back in 2024, we were seeing that engineers would get their first review within 3 hours…” — Will Bond, Uber

Thu, 15 Oct 2026 00:00:00 GMT

“Back in 2024, we were seeing that engineers would get their first review within 3 hours. Now in 2026, that has grown to 9 hours uh in addition to all of the volume changes. So, in short, code review is now the bottleneck that we are running into.” — Will Bond, Uber

Uber has thousands of engineers across hundreds of teams working in six language-specific monorepos. As PR volume and size grew, the time to a first review tripled, which is why Uber built uReview, its in-house multi-agent code reviewer. Quoted from YouTube’s auto-generated captions.

“This is in many ways the most painful part. Like I expect that people are a little bit…” — Zach Lloyd, Warp

Thu, 15 Oct 2026 00:00:00 GMT

“This is in many ways the most painful part. Like I expect that people are a little bit tired of reviewing a gentic slop. Uh I would have an agent do code review first and then it becomes over time like a riskmanagement exercise of like when do you bring in humans to do code review.” — Zach Lloyd, Warp

Lloyd calls code review the most painful stage of an agent-driven delivery loop. His answer is to have an agent review first, and to treat the question of when humans step in as a risk-management decision that changes over time. Quoted from YouTube’s auto-generated captions.

“So all PRs that get contributed to warp go through an agent managed review process and it…” — Safia Abdalla, Warp

Thu, 15 Oct 2026 00:00:00 GMT

“So all PRs that get contributed to warp go through an agent managed review process and it goes through multiple iterations and we don't actually ping any of the human reviewers on our team until an agent has approved our PRs. Um, which helps manage the workload a lot for the team.” — Safia Abdalla, Warp

When Warp open-sourced its client it received thousands of PRs from hundreds of contributors. Agents triage incoming issues and review every PR first, so the human maintainers only see changes an agent has already approved. Quoted from YouTube’s auto-generated captions.

“You're going to have a sort of control plane for figuring out how work gets distributed…” — Zach Lloyd, Warp

Thu, 15 Oct 2026 00:00:00 GMT

“You're going to have a sort of control plane for figuring out how work gets distributed across your factory floor. You're going to have the actual place where the work happens. And so that's going to be cloud sandboxes.” — Zach Lloyd, Warp

Lloyd describes the architecture a software factory ends up with: a control plane that decides how incoming work is distributed among agents, and separate cloud sandboxes where the agents do the work. Quoted from YouTube’s auto-generated captions.

“if you had a code review agent uh and it was leaving comments and a senior engine…” — Zach Lloyd, Warp

Thu, 15 Oct 2026 00:00:00 GMT

“if you had a code review agent uh and it was leaving comments and a senior engine engineering or team was going and correcting those comments, you'd want an observer agent that would look at that and basically uh improve the code review agent for the next run.” — Zach Lloyd, Warp

Lloyd describes a loop in which observer agents watch how factory agents perform. When engineers correct a review agent’s comments, the observer turns those corrections into improvements for the review agent’s next run. Quoted from YouTube’s auto-generated captions.

“Your [Sentry] your agent might get lucky and like be able to figure out the same root…” — Suraj Gupta, Warp

Thu, 15 Oct 2026 00:00:00 GMT

“Your [Sentry] your agent might get lucky and like be able to figure out the same root cause that it did the first time around, but that's not guaranteed. And and even if it does, you're you've probably wasted a lot of tokens re-gathering context on an issue that you've already fixed before.” — Suraj Gupta, Warp

In Warp’s Oz cloud agent platform, an outer-loop agent extracts facts from each run into a versioned memory store attached to the agent, and later runs report which memories they used. Humans can review, edit or delete memories, and skill updates land as pull requests a person reviews. Quoted from YouTube’s auto-generated captions.

“Every agent shares one running base environment and overrides only the service it’s…” — Michel Chatmajian, Abnormal AI

Wed, 14 Oct 2026 00:00:00 GMT

“Every agent shares one running base environment and overrides only the service it’s touching.” — Michel Chatmajian, Abnormal AI

At 200+ agent PRs a day, a full environment per branch was too expensive. Each testbox overrides only the service under change and gets private copies of the datastores it declares.

“A redirect only works if every consumer of the config layer honors it, and the ones that…” — Michel Chatmajian, Abnormal AI

Wed, 14 Oct 2026 00:00:00 GMT

“A redirect only works if every consumer of the config layer honors it, and the ones that don’t fail silently, because a working connection to the wrong database looks exactly like a working connection. We shipped two consumers that quietly ignored ours.” — Michel Chatmajian, Abnormal AI

Isolation that leaks fails silently. They added a capture agent that checks the connections each test actually makes against the service’s declared manifest.

agent-messenger

Wed, 14 Oct 2026 00:00:00 GMT

An MCP server for agent-to-agent messages and threaded conversations, backed by Beads.

“The step-based approach also made it simple to run migrations on hundreds of files…” — Charles Covey-Brandt, Airbnb

Wed, 14 Oct 2026 00:00:00 GMT

“The step-based approach also made it simple to run migrations on hundreds of files concurrently” — Charles Covey-Brandt, Airbnb

Planning that removes the coordination problem: each test file is its own unit with its own validation gates, so nothing shares state. 75% of target files migrated in four hours, though some long-tail files needed 50 to 100 retries.

“agents conflict. They step on each other’s changes. They duplicate work. They make…” — Luke Alvoeiro, Factory

Wed, 14 Oct 2026 00:00:00 GMT

“agents conflict. They step on each other’s changes. They duplicate work. They make inconsistent architectural decisions. And so the coordination overhead ends up eating up the speed gains all the while you’re burning tokens.” — Luke Alvoeiro, Factory

Factory’s Missions run features one after another and parallelize only read-only work, such as research and code review. Quoted from YouTube’s auto-generated captions.

“For instance, most coding tasks involve fewer truly parallelizable tasks than research…” — Jeremy Hadfield, Barry Zhang and colleagues, Anthropic

Wed, 14 Oct 2026 00:00:00 GMT

“For instance, most coding tasks involve fewer truly parallelizable tasks than research, and LLM agents are not yet great at coordinating and delegating to other agents in real time.” — Jeremy Hadfield, Barry Zhang and colleagues, Anthropic

“And also coordination debts. This is when you get really hairy merge conflicts because…” — Maggie Appleton, GitHub

Wed, 14 Oct 2026 00:00:00 GMT

“And also coordination debts. This is when you get really hairy merge conflicts because agents will touch in the same files or developers even doing duplicated work because they both picked up a thing and tried to finish it in one day.” — Maggie Appleton, GitHub

Her other line from the talk: “Agreeing on what to build is the new bottleneck.” Quoted from YouTube’s auto-generated captions.

“And while developers can now generate code and PRs faster than ever, we still only have a…” — Taroon Mandhana, Atlassian

Wed, 14 Oct 2026 00:00:00 GMT

“And while developers can now generate code and PRs faster than ever, we still only have a fixed number of senior engineers able to review them.” — Taroon Mandhana, Atlassian

Median time from PR to merge had passed three days, with an average 18-hour wait for the first review comment. The fix was an AI first reviewer on every PR.

BeadBoard

Wed, 14 Oct 2026 00:00:00 GMT

A live dashboard, task reservations and typed agent-to-agent mail, built on Beads.

Its README says it was inspired by Gas Town. The bb-pi orchestrator is marked as under construction.

“Two workers polling within the same window can both observe the same ready bead, both…” — amshancodes, on GitHub

Wed, 14 Oct 2026 00:00:00 GMT

“Two workers polling within the same window can both observe the same ready bead, both claim it, and both begin work. Result: duplicated work, near-identical PRs, wasted compute.” — amshancodes, on GitHub

Reported from an overnight Gas Town convoy, and closed the same day. A closed issue doesn't show that every variant of the race is gone.

“Multi-worker fleets polling with --claim currently double-fill.” — constellation-murdarch, on GitHub

Wed, 14 Oct 2026 00:00:00 GMT

“Multi-worker fleets polling with --claim currently double-fill.” — constellation-murdarch, on GitHub

Reported against bd 1.2.2 four months after #3570 closed, and still open in October 2026. The same kind of failure, but nobody has shown it is the same code path.