Here's a hands-on look at coral-code and the practical benefits that teams working on large codebases can expect.
Anyone these days can take a weekend project from 0 to MVP. A few can go from MVP to first 10 customers. Even fewer can reliably maintain a production codebase past the first 100 customers.
coral-code is for the codebases that can't afford to have thousands of lines of unsupervised code pushed to prod every day. We built it with professional engineering teams in mind.
The silent majority who watch the AI hype come and go, and are still using these tools conservatively because the cost of slop is not just ugly code, but instead something with real user data and monetary value at stake.
If you're optimistic about AI but take the risks of AI-generated code in production seriously, install coral-code and join our first cohort of early adopters.
https://t.co/lY8B1H4JRy
Coral Discovery explores how communities of AI agents can tackle optimisation problems by pursuing and improving different approaches in parallel. https://t.co/NHiCu41IO4
Discovery is an autonomous research system that puts a team of AI agents to work improving an existing model or algorithm.
The agents explore different approaches, test their changes and build on one another’s findings.
Across three simulated public reservoir benchmarks, Discovery improved calibration, forecasting and field development:
1. Reached the noise floor on a 13-parameter calibration task in about three hours.
2. Reduced forecast error from 0.863 to 0.274 in seven hours, a 68% improvement.
3. Raised a development plan’s modeled expected net present value from $1.255 billion to $1.881 billion, improving the outcome across all six held-out geological scenarios.
The gains came from specific engineering choices. A library of prior simulations gave calibration a stronger starting point.
Discovery makes that search process repeatable. Start with a working method, define the objective and constraints, then let agents investigate improvements in parallel.
Financial-crime teams work with more signals than investigators can review. A useful research system helps turn that volume into a better-ranked queue and more accurate estimates.
In these Coral Discovery experiments, four agents explored different approaches to two public-data tasks. Estimating reported venue revenue and ranking healthcare billing records against a later-exclusion label.
They tested executable candidates, combined useful ideas, and moved into new research directions as the search progressed.
The recorded revenue-estimation error fell from 1.308 to 0.565 RMSLE. In the audit-ranking experiment, recorded ROC AUC rose from 0.708 to 0.893. These are evaluator results on public reported-receipts and exclusion-proxy data.
The revenue search improved by changing how it used the evidence. Agents began with text, engineered-feature and neural models, then combined their outputs.
A later approach estimated each venue from similar labelled venues and learned an adjustment, bringing the search to its best score. In audit ranking, aggregating six years of provider history exposed changes that a single annual record leaves out; combining that history with geographic context improved the ranking further.
For a team building an investigation product, this is a practical role for autonomous research. Searching across modelling approaches while engineers define the prediction task and investigators decide how to use its output.
Some of the hardest engineering problems have a clear score but no known best answer. Faster code, cheaper data transfers, more reliable pathfinding.
AI agents can search for better solutions. But a team can waste its advantage if everyone follows the same leading idea.
Coral Discovery gives agents a shared direction board: a way to track different approaches and reward progress within each one.
Promising alternatives can keep developing, even before they beat the overall leader.
In our paper’s controlled experiments, using the same model and budget, Coral Discovery beat the reproduced MIT CORAL baselines on six of nine tasks. On the kernel engineering benchmark, Coral Discovery reduced execution from MIT CORAL’s published 1,103 cycles to 1,095. It surpassed that published result by evaluation no. 275. Fewer than half the 596 evaluations used in MIT’s reported run.
The opportunity is to make a team’s search broader while preserving its ability to build on what works.
Watch how Coral Discovery turns parallel effort into a more deliberate division of research.
Read the paper: https://t.co/L5K5ew4XTZ
Subagents help divide up a task. Coral’s agent graph lets agents working across a codebase stay aware of each other’s findings, share context, and follow connections as they emerge.
This demo takes a closer look at how that coordination works beyond the initial delegation.
Already a Codex user?
Use your subscription to power Coral Code's agent graph with GPT-5.6 Luna. https://t.co/nniO8S1x4i
Stay tuned this week for a deeper dive into how the agent graph works and how your harness interacts with it to improve its code certainty.
Thinking effort is only one part of the equation.
Coral Code gives your harnesses extra controls over permission levels and certainty, increasing the bandwidth between the codebase, agent, and developer.
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
We've expanded access modes on Coral Code so agents stop deleting your production databases.
PSI editing modes offer a middle ground between needing to approve every action and letting an agent run across your whole machine.
Try it here: https://t.co/nniO8S1x4i
Coral Code is now available on the JetBrains Marketplace.
AI code quality is solved.
While most of the industry has raced toward faster automation, longer running agent tasks, and bigger swarms of agents, we decided to tackle the basics of turning AI from a code generator to a reliable teammate.
Real production codebases with real stakes, serving real customers, are built through decades of human expertise.
We asked how we could turn that expertise into something agents can inherit.
Coral Code coordinates the harnesses you already use: Codex, Junie, and Claude Code, and gives them a graph of agents that they use to construct context dynamically and preserve the code conventions and quality standards your team has already built.
https://t.co/IOcFw9GeEZ
Many developers still think of context engineering as maintaining a list of markdown files.
This is one of the first things that breaks down at scale.
Coral Code gives Junie, Codex, and Claude a graph of agents that it can use to retrieve nuanced context across your codebase, surpassing even frontier models like Fable 5 in context accuracy.
Check out our research preview, available in JetBrains with future plugin expansion and a standalone runtime coming soon: https://t.co/lY8B1H4JRy
As code becomes commoditised, the challenge shifts from writing to understanding.
Coral Code layers an agent graph on top of your codebase, with each agent in the graph assigned focused responsibility across different regions.
From there, Codex, Junie, OpenCode, or whatever harness you already work with every day, gain a better understanding of your code and your engineering team's standards across long-horizon tasks.
Try it out! https://t.co/lY8B1H4c20
ill for sure be playing around with coordinated multi-agent coding on claude code
but I'll also be doing it on codex, junie, and opencode
ICYMI, our meta-harness generalises agent graph messaging across most of your favourite agents
give it a spin! https://t.co/LZ0gTtgRVj
We're excited to see agent-to-agent messaging becoming mainstream.
Our research has already shown that coordinated instances of Opus 4.6 can outperform a single instance of 4.8 in coding task accuracy.
This recent publication by VentureBeat covers how our messaging layer, and by extension, coral-code, our coding agent meta-harness, generalises these capabilities across other widely used harnesses such as Codex, Junie, and OpenCode.
https://t.co/CjpiUqVMSK
New in Claude Code: your sessions can now message each other.
Instead of having to re-explain yourself in another session, you can now tell Claude to do it. It sends a summary (not your history or files), and the other session picks it up mid-task.
Last year, we topped the GAIA benchmark with a team of GPT-4.1 mini agents, outperforming Microsoft's multi-agent framework by a wide margin.
Earlier this week, we outperformed Fable, Opus 4.8, and GPT-5.6 with a graph of DeepSeek v4 agents running on Coral Code.
Our stance has always been that frontier intelligence can be achieved horizontally, scaling via coordination of small and large models alike.
Andrej Karpathy just challenged one of the biggest assumptions in AI.
Most people think better agents come from bigger models.
He argues the opposite.
In 16 minutes, he explains why small models, the right tools, and closed feedback loops are where real capability comes from.
Watch it. Bookmark it.
Frontier coding agents are more capable than ever, but developers still spend a big portion of their time on context engineering.
Coral Code fixes that.
Join our earliest cohort and shape the future of coding intelligence.
https://t.co/tQYZRqHq2K
Coral Code makes models think smarter.
Join the earliest alpha testing group and see how multi-agent responsibility graphs can get you the most out of your GPT 5.6 usage.
https://t.co/tQYZRqHq2K
This 10-minute workshop will take you through everything you need to know to start building multi-agent systems with Coral using the Codex and Claude Code skills.
Topics covered include:
- Skill installation and prompting
- The 3 core building blocks for agent communication (threads, messages, mentions)
- Turning a sequential system into a dynamic agent group
- Blackboard and negotiation design pattern
- Building custom Codex CLI tools
- Using Coral Skills for long-horizon /goal tasks
Sign up for Coral Cloud to receive $30 in Coral LLM Proxy credits and follow along in your coding harness of choice.
Coral Cloud: https://t.co/roBN7PDUb7
Coral Skills: https://t.co/3byGmPhpbz