Active Research Infrastructure

Heterogeneous AI
Orchestration at
Production Scale

A runtime for composing task execution graphs across distributed, heterogeneous AI agents. Subtasks are routed by capability profile, context isolation is enforced at the session boundary, and validated outputs are synthesized through cross-agent consensus.

10⁴+
Addressable agent endpoints / logical worker nodes
<40ms
Median control-plane dispatch latency
99.7%
Validator quorum convergence on validation-eligible tasks
O(log n)
Indexed candidate retrieval / incremental placement complexity
System Architecture

Built on first principles

Every layer is designed around a single constraint: maximize effective throughput across heterogeneous inference capacity without sacrificing output quality.

01 — Agent Registry

Online Capability State

Per-agent n-dimensional capability state is maintained and updated online using task outcomes, validator verdicts, and runtime telemetry. The router searches over a continuously-calibrated capability manifold with no static tags and no manual configuration.

LISPEX
(define (update-state c-a alpha task)
  (let ((r (task-outcome task))
        (v (validator-verdict task))
        (t (runtime-telemetry task)))
    (v+ (vscale (- 1 alpha) c-a)
        (vscale alpha (psi r v t)))))
02 — Task Planner

Typed DAG Decomposition

Requests are decomposed into typed execution DAGs where each node carries I/O schema, QoS constraints, and privacy labels. The planner encodes both data dependencies and parallelizable branches, emitting a fully-typed graph consumed by the scheduler.

LISPEX
(define (make-dag vertices edges types qos)
  (list vertices edges types qos))
03 — Scheduler

Latency-Aware Constrained Dispatch

The scheduler retrieves candidates from the capability index in O(log n) time, then solves a constrained assignment over the bounded candidate set to minimize a weighted combination of latency, cost, and risk while satisfying capability thresholds and agent capacity limits.

LISPEX
(define (dispatch tasks agents)
  (letrec ((score
             (lambda (assignment)
               (+ (sum-lat  assignment)
                  (* lambda-c (sum-cost assignment))
                  (* lambda-r (sum-risk assignment)))))
           (feasible?
             (lambda (assignment)
               (and (covers-all?          assignment tasks)
                    (respects-capability? assignment)
                    (respects-capacity?   assignment)))))
    (argmin score
            (filter feasible?
                    (candidate-assignments tasks agents)))))
04 — Context Manager

Minimal-Sufficient Context Projection

Each agent receives the least-privilege context view its subtask requires. Windows are materialized by computing the dependency closure over the session graph, intersecting with per-agent policy filters and subtask schema. Cross-contamination is prevented without sacrificing coherence.

LISPEX
(define (context agent subtask)
  (redact
    (intersect
      (closure session-graph
               (ancestors subtask))
      (policy  agent)
      (schema  subtask))))
05 — Validation Layer

Validator Quorum & Adjudication

Critical outputs are routed to an independent validator set. Semantic similarity detects divergence; final accept/reject is determined by a composite score over reference grounding, hard constraint checks, and executable validators. Resolution is by quorum, not majority vote.

LISPEX
(define (score output)
  (+ (* w-g (ground   output R))
     (* w-h (hardcheck output C))
     (* w-e (exec     output T))))

(define (accept? outputs)
  (>= (length
        (filter (lambda (o)
                  (>= (score o) tau))
                outputs))
      quorum))
06 — Synthesis Engine

Typed IR Aggregation

Validated outputs are normalized into a canonical typed intermediate representation, then folded deterministically in topological order along the original DAG. A final schema validator gates the output, ensuring structural correctness before delivery.

LISPEX
(define output
  (validate-schema
    (fold-topological
      (map normalize outputs)
      dag)))
01
Task Input
Structured request + constraints
02
DAG Construction
Dependency analysis + parallelism
03
Agent Dispatch
Capability-matched assignment
04
Parallel Execution
Context-isolated inference
05
Validation
Cross-agent consensus check
06
Synthesis
Schema-conformant output
Capabilities

What this enables

Concrete outcomes made possible by coordinating heterogeneous inference at the infrastructure layer.

Compute Arbitrage

Workloads are routed to the minimum-effective-cost path satisfying SLO constraints, factoring in real-time capability state, latency, reliability, and unit cost. As frontier model pricing shifts, the scheduler adapts automatically.

Capability Composition

Specialized agents for code generation, formal reasoning, and multimodal understanding can be composed into a single task workflow. No agent needs to be a generalist; each contributes its strongest capability class.

Fault-Tolerant Execution

Agent failures trigger automatic subtask reassignment via ranked fallback candidates, checkpointed state, and idempotent replay semantics. Forward progress is maintained on SLA-bound tasks without side-effect duplication.

Context Efficiency

Minimal-sufficient context projection eliminates redundant token injection. Each agent processes only what its subtask requires, reducing per-request cost and enabling maximal parallelism across context-independent branches.

Auditable Provenance

Every output is attached to an immutable lineage record spanning agent identity, capability snapshot, context slice, validator verdict, and synthesis path. This supports debugging, compliance review, and capability improvement feedback loops.

Model-Agnostic Integration

A canonical invocation envelope and adapter contract allows frontier APIs, local runtimes, fine-tuned models, and specialist executors to register as first-class agent nodes with no platform changes required.

Research Direction

At production latency, the bottleneck is coordination, not capability

Raw model capability has scaled dramatically. The dominant unsolved problem is orchestration under constraints. Open-ended tasks must be decomposed into subtasks that can be reliably matched to agents, executed in parallel, and synthesized into coherent outputs at production latency, without human review in the loop.

Project Vercel is our long-term research infrastructure for this problem. We study online capability state calibration, constrained placement under SLO objectives, minimal-sufficient context materialization, and reference-anchored validator quorum, working toward a general-purpose coordination runtime for heterogeneous AI systems.

Capability Measurement
Benchmark-free online calibration of per-agent capability states. Vectors are updated continuously from task outcomes, validator verdicts, and telemetry. No held-out test sets required.
Graph Optimization
Constrained approximation for dynamic subtask placement under latency, cost, and SLO objectives. Current focus: bounded-ratio greedy placement with incremental rebalancing.
Context Compression
Learned minimal-sufficient context views under dependency closure. Determines the smallest context projection that preserves cross-subtask coherence, reducing token cost without quality degradation.
Consensus Mechanisms
Reference-anchored validator quorum and adjudication for distributed inference outputs. Combines semantic divergence detection with hard constraint checks and executable validators.
Deployment Context

Built independently, deployed proudly on Vercel

Project Vercel is an independent research initiative. It is not affiliated with or endorsed by Vercel, Inc. The name reflects our alignment with the broader ethos of fast, composable, production-grade infrastructure. The system itself runs on Vercel-powered deployment primitives.

Independent Project

Project Vercel is an independent initiative by Studio Haze. It is not an official Vercel product, subsidiary, or regional branch.

Built on the Vercel Stack

The system is built around the Vercel deployment platform and benefits from its deployment model, delivery ergonomics, and operational simplicity.

Name with Intent

The name reflects an affinity for the culture of fast iteration, clean deployment surfaces, and composable production systems. It signals admiration for that ecosystem, not formal affiliation.

Vercel is a trademark of Vercel, Inc. Project Vercel is an independent project and is not affiliated with, endorsed by, or sponsored by Vercel, Inc.

Contact

Work with us

We're exploring partnerships with research institutions, AI infrastructure teams, and organizations running demanding multi-agent workloads. If this work is relevant to your stack, we'd like to hear from you. We're especially interested in teams building demanding AI systems on modern deployment platforms and composable infrastructure stacks.

contact@vercel.co.kr