The factory

Seven coordinated stations, one learned orchestrator, one verification gate before anything ships.

Outputs that fail at Verify cycle back to Implement. Anything that passes ships as (code, spec, proof).

Each station is an ensemble of small models trained for that station's task, backed by structured external memory (AST graphs, proof DAGs, accumulated Miss→Rule libraries) offloaded across DRAM, NVMe, and computational storage. A learned orchestrator routes tasks, allocates context budget, and decides when to verify.

Why small models

Modern LLMs are weak at exploration and proof construction. These are exactly the operations software verification requires: walking adversarial paths, finding edge cases, constructing or checking invariants. A frontier model run once against a real codebase will miss most of what matters and produce confident-looking but unsound reasoning on the rest.

The fix is parallelism. A few hundred specialized agents probing the system from different angles will collectively find issues that any single agent would miss: fuzzers, invariant checkers, symbolic executors, exploit constructors, taint trackers, regression generators. Coverage produces the signal.

But running 200 frontier-model instances against one codebase costs more than a senior engineer's annual salary in a day. At current pricing, parallel verification at production density is economically impossible. This is why nobody runs it today.

The architecture we use is many narrow specialized small models, one per station, with a single frontier model handling the steps that genuinely require reasoning. We train these models on data the factory generates as it runs. Every customer workflow improves the next generation.

Inference

The factory runs in the background. We built an inference stack for exactly this profile: background, multi-agent, throughput-first.

Inference runs roughly 10–100× cheaper than a frontier-only baseline, and the same GPU hour that serves one chat response in an interactive system serves hundreds of agent steps in ours. The mechanics: GPU scheduling that mixes small-model calls with occasional frontier reasoning steps, aggressive use of spot capacity for the underlying hardware, and agents that self-evict their KV cache on preemption so spot interruptions never lose in-flight work.

At these costs, running hundreds of specialized agents against one codebase becomes affordable.

Vulnerabilities found

The same architecture, applied to the security station, has produced dozens of Critical and high-severity findings across production browsers, crypto infrastructure, databases, and AI systems, including multiple sandbox escapes in Chromium.

Publications

• Synthesizing Multi-Agent Harnesses for Vulnerability Discovery (AgentFlow). Formalizes harness synthesis for security agents; the foundation of the architecture described above.
• Your Agent Is Mine: Measuring Malicious Intermediary Attacks on the LLM Supply Chain (MINE). First work to characterize LLM routing as a supply-chain attack surface; the basis for the integrity layer in our production router.
• Semia: Auditing Agent Skills via Constraint-Guided Representation Synthesis (SEMIA). Audits hybrid skill specifications as Datalog before deployment.

Team

AgentSey is built by researchers in programming languages, formal verification, and computer security, with operating experience in AI infrastructure. Members of the team have received Distinguished or Best Paper Awards at PLDI 2018, PLDI 2022, ASE 2020, CHI 2021, and ACM SIGSOFT, alongside a Google Research Scholar Award and Ethereum Academic Award. Earlier work from the group is widely cited as foundational in formal methods and program analysis.

Between us we have shipped a certified JavaScript compiler verified end-to-end, a multi-agent security system responsible for dozens of Critical and high-severity findings in production systems, the only production LLM router with a research-grade integrity layer, a managed cloud runtime for autonomous agents, and an in-house training stack for the specialized small models behind the factory.

We are hiring

We are hiring security researchers, systems engineers, and ML specialists to extend the factory and the infrastructure beneath it. The work spans program analysis, fuzzing, exploitation, sandboxing, formal verification, large-scale agent orchestration, small-model training, and storage-aware inference. The best candidates move comfortably between low-level systems work and agent system design.

For careers, research collaborations, and pilot inquiries, contact hr@riema.xyz.