July 29th news: Today, AI expert, former vice president of Alibaba Cloud, and founder of Lepton AI...Jia YangqingAbout a month after leaving Nvidia, he started another business, announcing the launch of Intent Lab and establishing an autonomous AI team called Fleet, hoping to...Transform user-defined intents directly into production-grade software systems..
▲ Jia Yangqing introduces Intent Lab's products (Image source: X)
Unlike programming agents that are primarily responsible for writing code, Fleet attempts to take on the work of a complete engineering team, from understanding requirements, designing architecture, and coordinating tasks, to writing code, testing and verifying, and continuous improvement after deployment.
Intent Lab announced three early results from Fleet:
First, by redesigning the TensorRT-LLM code, a large model inference optimization tool, the inference speed of GLM-5.2 was improved.Increased by 534%Intent Lab states that this is currently...GLM-5.2's fastest inference speed.
Second, based on an initial instruction, a database matching the external behavior of SQLite is built from scratch, and the entire round of execution generates...87.2 million output tokensUltimately passedAll 6 million acceptance testsWhen using all open-source models, the cost is approximately$350 (approximately RMB 2368).
Third, we built AgentFS, a distributed file system for agents, achieving performance levels comparable to EFS in some aspects.626 timesIt also completed formal model checks on approximately 1.9 million states and approximately 300 integration tests.
01.
From understanding intent to continuous evolution
Covering all six stages of software development
Intent Lab states that Fleet works like a software team that follows engineering principles, and the entire development process includes...Understanding, designing, coordinating, building, validating, and evolvingSix steps.
Since most projects begin with a rather vague intention, Fleet will first...Define goals together with usersThis involves translating the findings into concrete results and defining constraints and acceptance criteria. All subsequent work will revolve around these criteria.
After understanding the requirements, Fleet weighs different solutions, determines the system interfaces and architecture that meet the requirements, and continuously records how each decision affects other parts of the system. For large systems composed of multiple components, it also breaks down tasks, assigns responsibilities, and manages the dependencies between tasks.
Fleet will continue to implement the system while refining the design. As new information emerges during development, it will update the architecture and code accordingly to ensure the design remains consistent with the actual running system.Validation will also be carried out throughout the entire build process.Fleet uses a combination of formal proofs, integration tests, and runtime fault injection to detect problems as early as possible while the code is still changing.
Fleet's work doesn't stop once the software is deployed in production. It continues to monitor the system's actual performance and feeds back information such as usage, performance, and cost into the design, enabling the software to continuously improve as user needs evolve.
02.
GLM-5.2 inference speed increased by 534%
Reaching 647 tokens/s
The first task Intent Lab assigned to Fleet was to redesign the code of NVIDIA's large model inference optimization tool TensorRT-LLM, enabling GLM-5.2 to run on Grace Blackwell's compute nodes, and to autonomously find, implement, and verify optimization solutions.
Since TRT-LLM has already undergone extensive performance optimization, further speed improvements are not easy. Fleet ultimately...Kernel, runtime, communication and speculative decodingWe can approach this from four aspects.
kernelAt the kernel level, Fleet reduces kernel startup and improves inference performance by using kernel fusion and agent-generated PTX and SASS code.twenty four%;existruntimeAt the level of H2D batch processing, steady-state decoding with zero copy is achieved, thereby improving performance.16%;communicationAt the layer level, aggregated communication, residual addition, and RMSnorm are integrated to improve performance.18%;existSpeculation DecodingIn terms of this, by adding DSpark, multiple candidate tokens can be proposed and verified at once, further improving inference speed by approximately [missing information].4 timesUltimately, end-to-end inference performance is improved compared to the original TRT-LLM.534%.
▲Performance improvements in kernel, runtime, communication, and speculative decoding (Source: Intent Lab)
The entire optimization process is driven autonomously by Fleet. The system first analyzes the performance limits and identifies bottlenecks, then proposes and verifies optimization solutions. Solutions that fail verification are abandoned and retried, while verified solutions are retained and proceed to the next round of optimization.
Test results show that the original TensorRT-LLM runs GLM-5.2 at a speed of [speed value missing].102 tokens/sAfter kernel, runtime, and communication optimizations, it was upgraded to161 tokens/sAfter joining DSpark, it reached647 tokens/send-to-end performanceIncreased by 534%.
▲ Output speed improvement (Image source: Intent Lab)
According to Intent Lab, this is also the fastest inference speed achieved by GLM-5.2 to date. The entire process is completed end-to-end by Fleet without human intervention.
03.
Build the database from scratch with a single instruction
87.2 million output tokens were generated.
Fleet's second task isBuild a new system that is compatible with the existing system but whose internal architecture is completely redesigned from scratch..
System modernization is not simply about modifying old code; it's about redesigning the internal architecture while maintaining external behavior. In this task, Fleet did not refer to the existing SQLite code and documentation, but instead used the SQLite test suite as the acceptance criterion.
Intent Lab provides only one instruction: "Build a SQL database engine that is compatible with SQLite and can pass all sqllogictest test cases with performance equal to or exceeding that of SQLite."
Fleet assigns tasks to different roles, including decision-making, architecture design, coding, testing, review, and quality assurance. These roles collaborate repeatedly throughout the process, continuously adjusting the design, conducting tests, and reviewing the code as it changes.
The entire round of operation generated a total of87.2 million output tokens.in,Coding accounted for 38.6%, testing for 21.7%, decision-making for 17.8%, and architecture design for 13.3%.Review and quality assurance together accounted for 8.6%. This means that building a database is not just about generating code; a significant amount of work is also involved in design, decision-making, and validation.
▲Distribution of Fleet's output tokens (Source: Intent Lab)
Fleet modernized the database based on a single line of prompts, designed and validated the system from scratch, and ultimately passed...All 6 million acceptance tests.
The same Fleet can be used with different models to complete the task. When using Opus 4.8, the total runtime cost is approximately...$2,000 (approximately RMB 13,532)When using the entire open-source model, the cost is approximately...$350 (approximately RMB 2368).
04.
Building a distributed file system from scratch
Directory search speed reaches 626 times
Fleet's third task is to build a distributed file system in the cloud environment optimized for Agent usage and to rigorously check its correctness and fault tolerance.
Agents frequently create sandboxes, scan directories, and process a large number of small files in shared cloud storage. Intent Lab tried off-the-shelf solutions such as Amazon EFS and S3FS, but these products all had limitations under Agent workloads, so Fleet built AgentFS from scratch.
In the mdtest test, AgentFS performed directory and file creation operations faster than EFS.45 timesThe file read speed is EFS.30 timesIts advantages are even more pronounced in the status query operations frequently used by agents: directory status query speed reaches EFS speed.626 timesFile status query speed reaches EFS625 times.
▲ mdtest test results (Image source: Intent Lab)
In Git repository tests, AgentFS cloning the repository was approximately as fast as EFS.2.5 to 3 timesThe speed of querying the code repository status is approximately6.9 times to 9.5 timesThe speed of deleting the code repository is approximately13 to 14 timesThe tests used etcd and Kubernetes as small and large codebases, respectively. S3FS failed to fully detect the Kubernetes codebase, thus only obtaining partial results.
▲ Git codebase test results (Image source: Intent Lab)
In addition to performance, AgentFS has undergone rigorous validation. Fleet examined approximately 1.9 million reachable states across 31 models, covering different concurrency sequences, system crashes, and race conditions. AgentFS also completed approximately 300 integration tests, as well as fault injection and fuzzing tests.
During the verification process, Fleet discovered a bug introduced by the programming agent that could lead to data corruption during distributed creation and deletion operations. After fixing the bug and re-verifying, the corrupted state in the system was eliminated.
05.
Conclusion: From "AI Writing Code" to "AI Building Systems"
Intent Lab is experimenting with assembling multiple agents into an autonomous engineering team with different roles. From inference engine optimization and database reconstruction to the creation of a distributed file system, Fleet needs to design solutions, coordinate tasks, complete development, and verify results. This means that AI programming is attempting to move beyond simple code generation and enter a complete software engineering process.
However, these are still early results published by Intent Lab. Whether the autonomous AI team can stably handle more complex real-world projects and operate at an acceptable cost over the long term requires further practical verification.
Source: Intent Lab, X