This example shows 400+ spawned agents, but not in the way you typically see agent swarms of this size.
I've seen a lot of swarm-like agent systems that end up doing less with more.
"You are a backend engineer ..."
"You are a security specialist ..."
"You are an expert code reviewer"
Frankly, role-based subagents are mostly dead and end up producing slop.
Coral Code uses a graph of agents to make your primary agent (Codex, Claude, Junie) passively aware of evidence relevant to the request as it's working.
Each agent in the graph is not given an anthropomorphic role, instead they are assigned responsibility for a certain file or specific snippets of code within larger files.
This means that each agent's context window is used to maximum efficiency, holding information relevant to its area of responsibility for the primary agent to use.
passive awareness is the system powering agent graphs in Coral Code, and can be used to power virtually any generalised use of subagent or multi-agent systems
link to full paper: https://t.co/ihi7GJzbLa
runtime duration rarely correlates with correctness
some of the sloppiest code I've seen has come from multiple subagents running for 8/14/24+hrs
certainty is the new moat, and this is what Coral Code enables
see the difference yourself https://t.co/WQaqdijVcY
the codebase being worked on is the winning source of truth when it comes to conventions, the level of contextual detail the agent graph provides in itself helps keep any harness on track. also the way coral's communication tools work, its not really about coral per se imposing coordination rules, it just exposes tools for agents to communicate between themselves so disagreements are handled by one harness "pinging" another in real time
This is honestly a great example of why the average working class has such a low opinion of the “SF-class”
-> The SF-class makes gen AI mainstream
-> goes on roadshows talking about how its gonna replace half or more of all white collar knowledge work
-> white collar knowledge workers get force-fed AI
-> white collar knowledge workers begrudgingly adopt it, just say “fuck it, let me just automate everything with AI then”
-> and then the same SF-class comes back around with terms like “meat proxies” to tell those same disenfranchised knowledge workers that they’re “not using it good enough”
Don’t get me wrong, I’m radically against AI slowdowns, government intervention, and regulation. But I can empathise with the vast majority of workers who could care less about optimal AI use and are just using whatever tools they’re forced to adopt so they can keep their paycheck coming in, not everyone is in this for some grand vision of acceleration, especially not in this economy.
@AnkanXplorer vercel was already a go to choice way before AI takeover
whether you're doing everything manually or telling an agent to do it, simplicity of vercel lets you defer more complex infra ownership and maintenance until you've actually proven some level of traction in your idea
i dont think i can ever go back to sensitive migration work without coral-code
those disaster screenshots you see of an agent deleting entire production databases are probably a side effect of starting a new session and hoping the agent reconstructs context the right way
if the context your agent ingests for these types of tasks relies on docs you wrote 2 weeks ago, how well you explained your initial prompt, and other old school "context engineering" workflows, then you're probably cooked
the content creators (unless they're actually cracked engineers) telling you to run swarms of parallel coding agents are probably BSing you
ive been playing around a lot with codex's somewhat new thread messaging feature:
start in one planning session -> have it orchestrate tasks into new threads -> reconcile results, review, make changes if needed, or continue looping
ive found that more often than not, even if task orchestration looks solid at the start, context eventually breaks down when its in its second to third loop of dispatching tasks, reviewing, and dispatching again
the most consistent results I've gotten with multi-agent coding is having a single "outer agent" dynamically construct and change context it needs by communicating with a graph of "inner agents"
the contrast is that the outer agent isn't really an orchestrator. it's still responsible for implementation, but having agents as context sources lets it be more efficient with its own context window, makes it less reliant on stale docs that I'm lazy to update, and keeps execution legible
TLDR: multiple agents steering context, single agent controlling execution
@Hesamation there's a punchline in here about regulatory capture or about dario trying to kill open source for "safety" reasons but im too brainrotted to write it
@grok do your thing
i'm so glad we're past the "software eng is over" phase
i think it's worth reflecting on two kinds of idiots
- people who said it was over
- people who reacted by saying it wouldn't change
both of them missed out on a lot of fun!
the amount of inference power users on pro plans have been getting has always been disproportionately in their favour. oai is actually losing a lot of money for certain users so the math isn't as clean as "if a billion people paid $1"
not gonna pretend to know the exact numbers but a $200/month user can easily lose oai $2000+
the content creators (unless they're actually cracked engineers) telling you to run swarms of parallel coding agents are probably BSing you
ive been playing around a lot with codex's somewhat new thread messaging feature:
start in one planning session -> have it orchestrate tasks into new threads -> reconcile results, review, make changes if needed, or continue looping
ive found that more often than not, even if task orchestration looks solid at the start, context eventually breaks down when its in its second to third loop of dispatching tasks, reviewing, and dispatching again
the most consistent results I've gotten with multi-agent coding is having a single "outer agent" dynamically construct and change context it needs by communicating with a graph of "inner agents"
the contrast is that the outer agent isn't really an orchestrator. it's still responsible for implementation, but having agents as context sources lets it be more efficient with its own context window, makes it less reliant on stale docs that I'm lazy to update, and keeps execution legible
TLDR: multiple agents steering context, single agent controlling execution
@pukerrainbrow that's the beauty of the free market
one jurisdiction tries to slow down innovation, another jurisdiction complies, and another one just doesn't and "wins by doing nothing"