Agent orchestration for the timid
substack.com
substack.com
My (boring b2b/b2e) org has scripts that wrap a small handful of agent calls to handle/automate our workflow. These have been incredibly valuable.
We still 'yolo' into PRs, use agents to improve code quality, do initial checks via gating. We're trying to get docs working through the same approach. We see huge value in automating and lightweight orchestration of agents, but other parts of the whole system are the bottleneck, so theres no real point in running more than a couple of agents concurrently - claude could already build a low quality version our entire backlog in a week.
Is anyone exploring the (imo more practically useful today) space of using agents to put together better changes vs "more commits"?
That’s what I’ve been focused on the last few weeks with my own agent orchestrator. The actual orchestration bit was the easy part but the key is to make it self improving via “workflow reviewer” agents that can create new reviewers specializing in catching a specific set of antipatterns, like swallowing errors. Unfortunately I've found that what decides acceptable code quality is very dependent on project, organization, and even module (tests vs internal utilities vs production services) so prompt instructions like "don't swallow errors or use unwrap" make one part of the code better while another gets worse, creating a conflict for the LLM.
The problem is that model eval was already the hardest part of using LLMs and evaluating agents is even harder if not practically impossible. The toy benchmarks the AI companies have been using are laughably inadequate.
So far the best I’ve got is “reimplement MINPACK from scratch using their test suite” which can take days and has to be manually evaluated.
Yes, I am, although not really in public yet. I use the pi harness, which is really easy to extend. I’m basically driving a deterministic state machine for each code ticket, which starts with refining a short ticket into a full problem description by interviewing me one question at a time, then converts that into a detailed plan with individual steps. Then it implements each step one by one using TDD, and each bit gets reviewed by an agent in a fresh context. So first tests are written, and they’re reviewed to ensure they completely cover the initial problem, and any problems are addressed. That goes round a loop till the review agent is happy, then it moves to implementation. Same thing, implementation is written, loop until the tests pass, then review and fix until the reviewer is happy. Each sub task gets its own commit. Then when all the tasks are done, there’s an overall review that I look at. Then if everyone is happy the commits get squashed and we move to manual testing. The agent comes up with a full list of manual tests to cover the change, sets up the test scenarios and tells me where to debug in the code while working through each test case so I understand what’s been implemented. So this is semi automated - I’m heavily involved at the initial refine stage, then I check the plan. The various implementation and review loops are mostly hands off, then I check the final review and do the manual testing obviously.
This is definitely much slower than something like Gas Town, but all the components are individually simple, the driver is a deterministic program, not an agent, and I end up carefully reviewing everything. The final code quality is very good. I generally have 2-4 changes like this ongoing at any one time in tmux sessions, and I just switch between them. At some point I might make a single dashboard with summaries of where the process is up to on each, and whether it needs my input, but right now I like the semi manual process.
A concrete example of this is our Adviser Skill experiment [2]. In most AI workflows, a "reviewer" agent just dumps markdown feedback. Our Adviser doesn't just "talk"; it outputs an AISP 5.1 document ( a kind of "Assembly Language for AI Cognition" )
This document forces the agent to define:
- Strict Type Definitions for the issues identified (e.g., distinguishing between a gap, an edge case, or a missing requirement).
- EARS Rules (Easy Approach to Requirements Syntax) that determine the verdict. For example, a rule might state: "If any issue has a severity of ⊘ (critical), then the workflow MUST halt."
- Formal Evidence: Every "approve" or "reject" verdict must include a confidence score (δ) and a grounding proof (π) that explains why the change matches the original specification.
By treating the agent's output as a proof-carrying protocol rather than just text, we can chain multiple specialized agents (Architect, Strategist, Auditor) who "triangulate" on the codebase. They must reach a formal consensus where the variance between their scores is low.
This shifts the agent's goal from "Finish the task at all costs" to "Prove that this change is safe and correct." It turns out that iterating on the verification logic is much more effective for building reliable systems than just increasing the number of agents running concurrently.
[1] Brad Ross AISP : https://github.com/bar181/aisp-open-core
[2] Adviser skill : https://github.com/pnocera/skilld
Mayybe for some things you could set it up so that the screen output is livestreamed back into the agent, but I highly doubt that anyone is doing that for agents like this yet
For host-side code the agent can throw in a bunch of logging statements and usually printf its way to success. For device-side code there isn't a good way to output debugging info into a textual format understandable by the agent. Graphical trace viewers are great for humans, not so great for AI right now.
On the other hand, Cline's harness can interact with my website and click on stuff until the bugs are gone.
Now they can be promoted from junior coders into mid-level coders :)
Seems Codex works just fine with nsys and sqlite3 available to it, I've had success using it for debugging CUDA code that was crashing, and also for optimizing code.
What do you mean by streaming? LLMs aren’t that advanced yet where they can consume a live video feed but people have been feeding them screenshots from Playwright and desktop apps for years (Anthropic even released the Computer Use feature based on this).
Gemini has the best visual intelligence but all three of the major models have supported this for a while. I don’t think it’d help with fixing subtle problems in shadows but it can fix other gui bugs using visual feedback.
In terms of the state of software quality, the bar has actually been _lowered_, in that even major user-facing bugs in operating systems are no longer a showstopper. So it's no surprise to me that people are vibe-coding things "in prod" that they actually sell to other people (some even theorize claude code itself is vibe-coded, hence its bugs. And yet that hasn't slowed down adoption because of the claude max lock in).
So maybe one alternate way to see the "productivity gains" from vibe-coding in deployed software is that it's actually a realization that quality doesn't matter. The seeds for this were already laid years back when QA vanished as a field.
LLMs occupy a new realm in the pareto frontier, the "slipshod expert". Usually humans grow from "sloppy incompetent newb" to the "prudent experienced dev". But now we have a strange situation where LLMs can write code (e.g. vectorized loops, cuda kernels) that could normally only be done by those with sufficient domain knowledge, and yet (ironically) it's not done with the attention and fastidiousness you'd expect from such an experienced dev.
Case in point - https://github.com/NVlabs/vibetensor/blob/main/docs/vibetens...
> VIBETENSOR is an open-source research system software stack for deep learning, generated by LLM-powered coding agents under high-level human guidance. In this paper, “fully generated” refers to code provenance: implementation changes were produced and applied as agent-proposed diffs; validation relied on builds, tests, and differential checks executed by the agent workflow, without per-change manual diff review.
It overwhelms everyone's ability to keep track of what it's doing. Some people are just no longer keeping track.
I have no idea if people are just doing this to toy projects, or real actual production things. I am getting the sneaking suspicion it's both at this point.
I haven't yet tried gas town (or any of the mentioned tools) as I don't need so many agents that I need something like that plus the cost concerns. I've been rolling my own very light orchestrator (mostly just worktrees/branches/instructions) and relying on claude itself to manage the sub agents as necessary.
I was a bit surprised by the "ripping out beads" sentence from all of the article, as beads does seem to serve a purpose independent of the orchestration tools. Giving agents a ticketing system independent of what us humans use makes a lot of sense to me.
I've experimented with using Jira/Linear to handle the "current work todos" and using beads just seems so much better. No mcps and remote api calls is pretty great.
I'll be curious to see how the other orchestration tools are handling this, because it seems like they will have to handle it.
I'll play with that a bit to see what happens.
It has Playbooks for repeatable workflows using which you can model both generic SDLC workflows (Plan-Code-Review-Security review-Merge) or complex workflows like Language migration, Tech stack migration (Problem breakdown-Plan-Migrate-IntegrationTest-TechStackExpertReview-CodeReview-Merge)
Hopefully, it will have the least amount of changes once Claude Swarm and Claude Tasks becomes mainstream
Orchestration is cool but a sane orchestration setup with VM's is where it's at.
I've been working on orchestration for the past little while and I've got a very tight loop going where everything is in worktrees and containerized, all services are isolated, so I can easily work on db schema/migration stuff while a few other agents do frontend work etc. Getting Conductor to play nice with vm's was very tricky as their docs say they have no intention of implementing vm's and wrote a "trust me bro, it won't erase your system" blurb about it in their docs [0]
[0] https://docs.conductor.build/faq#what-permissions-do-agents-...
Brutal. I had to renice it and do some shenanigans to make it work crispy but now it's decent. Has a lot of hang ups like if you leave a text input focused for too long the fans on your laptop will start.