CustomGPT Agent

An enterprise-grade AI agent powered by Claude, GPT, Gemini and CustomGPT.ai

---

🏆 Results — #1 on the GAIA Leaderboard

CustomGPT.ai Research Lab is #1 on the GAIA benchmark — the top-ranked general AI assistant agent in the world, ahead of every major lab.

Rank	Agent	Organisation	Average score
🥇 1	CustomGPT.ai Research Lab v44	CustomGPT.ai	93.36%
2	Co-Sight Pro v1.0.1	ZTE-AICloud	93.02%
3	OPS-Agentic-Search	Alibaba Cloud	92.36%
4	Lemon	LR AILab, Lenovo CTO Org	91.36%
5	Nemotron-ToolOrchestra-0106	NVIDIA	90.37%
6	SU Zero – Shuqian Series Pro MAX	Suzhou AI Lab / Shuqian Tech	90.03%
7	HALO V1217-1	Microsoft AI Asia – Ads	89.37%

_{Leading entries on the GAIA Leaderboard as of 5 June 2026. CustomGPT.ai holds the #1 position.}

---

Overview

CustomGPT Agent is a multi-agent AI system that combines CustomGPT.ai's enterprise RAG platform with frontier-model reasoning and secure sandboxed code execution. It is designed to solve complex, real-world tasks that require multi-step reasoning, web research, file analysis, visual browser navigation, and computation.

The agent ranks #1 on the GAIA Benchmark (General AI Assistants) — a challenging benchmark that tests AI systems on tasks requiring real-world tool use, multi-step reasoning, and web browsing — tasks that go far beyond simple question answering.

The orchestrator is provider-agnostic: it runs on Claude Opus 4.6/4.7 via the Claude Agent SDK or on OpenAI GPT-5.5 through a parallel orchestrator, with a configurable thinking level per run. For tasks that require seeing a page — Google Street View, 3D model viewers, interactive maps — the agent drives a real headed Chromium browser through a vision-first computer-use tool suite, with Google Gemini 3.1 Pro as a secondary vision provider. Every step is wrapped by a quality & verification plane that re-checks claims against primary sources before any answer is returned.

---

Architecture

The system is organized into four planes: a provider-agnostic orchestrator, a set of specialist subagents, a shared MCP tool layer, and a quality & verification plane that wraps every step.

┌─────────────────────────────────────────────────────────────────────────┐
│              ORCHESTRATOR  (provider-agnostic, ReAct loop)               │
│     Claude Opus 4.6 / 4.7 (Agent SDK)   •   OpenAI GPT-5.5               │
│     Configurable thinking level  •  Plan → Act → Verify                  │
│     Experience memory + runtime-error memory injected into the prompt    │
└──┬────────┬────────┬────────┬────────┬────────┬─────────────────────────┘
   │        │        │        │        │        │   delegates via Task tool
┌──▼───┐ ┌──▼───┐ ┌──▼────┐ ┌─▼─────┐ ┌▼──────┐ ┌▼───────┐
│PLAN- │ │LOOKUP│ │COMPUTA│ │ FILE  │ │VISUAL │ │ CRITIC │
│NER   │ │(Opus)│ │-TION  │ │ANALY- │ │ANALYST│ │ (Opus) │
│      │ │web/  │ │(Sonnet│ │SIS    │ │(headed│ │grounded│
│plan  │ │wiki  │ │double-│ │(Sonnet│ │browser│ │re-check│
│      │ │rsrch │ │compute│ │ vision│ │ cu_*) │ │vs src  │
└──┬───┘ └──┬───┘ └──┬────┘ └──┬────┘ └──┬────┘ └─┬──────┘
   │        │        │         │         │        │
┌──▼────────▼────────▼─────────▼─────────▼────────▼───────────────────────┐
│                       MCP TOOL LAYER  (~80 tools)                        │
│  Search/Browse:  web_search · search_wiki · search_google · query_arxiv │
│                  · ask_perplexity · jina_search · jina_read ·           │
│                  browse_webpage · search_archive · find_paper ·         │
│                  fetch_via_wayback                                       │
│  Media/Video:    youtube_metadata · youtube_captions · youtube_search · │
│                  identify_song · extract_video_frames · extract_audio   │
│  Execution:      execute_python · execute_shell · read_file · write_file│
│                  (E2B microVM, Playwright)                              │
│  Vision/CU:      describe_image · image_question · cu_screenshot ·       │
│                  cu_click · cu_type · cu_scroll · cu_zoom · cu_navigate  │
│  Knowledge:      search_knowledge_base  (CustomGPT.ai RAG)              │
│  Reasoning/Mem:  deep_think · fetch_experiences · submit_answer          │
└──┬──────────────────────────────────────────────────────────────────────┘
   │
┌──▼──────────────────────────────────────────────────────────────────────┐
│              QUALITY & VERIFICATION PLANE  (hooks + loops)               │
│  • Quadruple-Verification hooks — 4 cycles on every tool output          │
│      C1 code quality · C2 security · C3 stop-quality · C4 research       │
│  • Self-Correction loop — render → screenshot → analyze → fix (×3)       │
│  • Majority-vote self-consistency across N runs                          │
│  • Budget / loop-detection / audit-log / safety hooks                    │
└──────────────────────────────────────────────────────────────────────────┘

Components

Component	Technology	Role
Orchestrator	Claude Opus 4.6 / 4.7 (Agent SDK) or OpenAI GPT-5.5	Task decomposition, routing, evidence-backed reasoning, and final answer synthesis — model & thinking level selectable per run; prompt augmented with experience + runtime-error memory
PLANNER Subagent	Claude / GPT-5.5	Produces a structured plan (question restatement, strategy, micro-step queue) before execution begins
LOOKUP Specialist	Claude Opus + search/browse tools	Web/wiki/arXiv/Perplexity research and fact retrieval, page navigation
COMPUTATION Specialist	Claude Sonnet 4.6 + E2B sandbox	Python/shell execution, math, and data processing with double-compute verification
FILE ANALYSIS Specialist	Claude Sonnet 4.6 + E2B + vision	Document parsing (PDF, JATS XML, Excel, images, audio), transcription, and file manipulation
VISUAL ANALYST Subagent	Claude / Gemini 3.1 Pro + headed Chromium	Vision-first navigation of Street View, 3D viewers, and interactive maps via the cu_* computer-use tools
CRITIC Verifier	Claude Opus (real research tools)	Grounded verification — independently re-checks the answer's source quote against primary sources before approval
E2B Sandbox	Firecracker microVM + pooled warm starts	Isolated Python/shell execution, Playwright browser automation, file workspace

Quality & Verification Plane

The layer that wraps every step and is the main driver of the agent's accuracy gains:

Mechanism	What it does
Quadruple Verification	Claude Agent SDK hooks running four cycles on tool output — C1 code-quality, C2 security, C3 stop-quality, C4 research-claim verification — with an immutable audit log
Grounded CRITIC	Re-checks the answer's cited source against primary sources with real research tools, not self-reflection, before any answer is submitted
Self-Correction Loop	Create → render → screenshot → analyze → fix loop (max 3 iterations) that makes generated artifacts executive-ready
Self-Consistency Voting	Runs a task N times and selects the majority answer (normalized grouping), with early-stop on consensus
Experience & Runtime Memory	Deterministic matchers inject lessons from past successes (experiences.json) and past failures (runtime_memory.json) into the orchestrator prompt
Sandbox Pool	Keeps warm E2B microVMs to remove cold-start latency across parallel runs
Operational hooks	Per-agent / global tool-call budgets, loop detection, safety gates, and cost tracking

---

Key Features

1. Multi-Provider Orchestrator — The same task can be driven by Claude Opus 4.6/4.7 (via the Claude Agent SDK) or OpenAI GPT-5.5, with a per-run thinking-level knob, enabling apples-to-apples cross-model comparison on identical specialists

2. Plan → Act → Verify Loop — A dedicated PLANNER subagent writes a structured plan.md before execution; the orchestrator acts against it and the CRITIC verifies the result, with durable notes.md observations for human visibility

3. Vision-First Computer Use — A real headed Chromium browser (persistent cookie profile, 16:9 viewport) is driven through a cu_* tool suite, letting the agent see and operate Street View, 3D model viewers, and interactive maps exactly as a human would

4. Dual Vision Stack — Claude vision plus Google Gemini 3.1 Pro as a secondary provider for exhaustive image description and strict image-vs-image verification

5. First-Party Media Tooling — YouTube metadata/captions/channel search, song identification, and frame/audio extraction give the agent direct access to video content instead of brittle scraping fallbacks

6. Grounded CRITIC Verification — The CRITIC re-checks the answer's cited source against primary sources with real research tools before any answer is submitted

7. Quadruple-Verification Hooks — Four independent cycles (code quality, security, stop-quality, research-claim) run automatically on tool output, with an immutable audit trail

8. Self-Correction + Self-Consistency — A render → screenshot → analyze → fix loop produces executive-ready artifacts; majority-vote across N runs reduces variance on hard tasks

9. Experience & Runtime Memory — Learns across runs by injecting prior successes and prior failures into the orchestrator prompt

10. Enterprise RAG Integration — CustomGPT.ai's 40+ API endpoints provide citation-backed knowledge retrieval from proprietary data sources

11. Full Observability — A model-I/O logger captures every request/response to JSONL and a progress bus emits real-time task events for live monitoring and replay

---

How It Works

User Question
     │
     ▼
┌──────────────┐
│   PLANNER    │──── Writes plan.md: restate question, choose strategy, queue steps
└──────┬───────┘
       │
       ▼
┌──────────────┐
│ Orchestrator │──── Routes each step to the right specialist (Claude or GPT-5.5);
└──────┬───────┘     injects matched experience + runtime-error memory
       │
       ├──► LOOKUP:        Search web/wiki/arXiv/Perplexity, browse pages
       ├──► COMPUTATION:   Write & run Python in the E2B sandbox (double-compute)
       ├──► FILE ANALYSIS: Parse PDFs, spreadsheets, images, audio
       ├──► VISUAL ANALYST: Drive headed Chromium — screenshot, click, zoom, pan
       └──► CustomGPT KB:  Query the enterprise knowledge base
              │
              ▼   (every tool output passes the 4 verification cycles)
       ┌──────────────┐
       │    CRITIC     │──── Re-check cited source against primary sources
       └──────┬───────┘
              │
              ▼
       ┌──────────────────────┐
       │  Self-correction ×3   │──── Render → screenshot → analyze → fix
       └──────┬───────────────┘
              │
              ▼   (optional) majority vote across N runs
        Final Answer (with plan.md, notes.md & evidence trace)

---

About CustomGPT.ai

CustomGPT.ai is an enterprise AI platform that enables businesses to build custom AI agents powered by their own data — with no hallucinations and full source citations.

Key Capabilities:

• 1,400+ data formats — Ingest websites, documents, helpdesks, videos, and more
• Anti-hallucination technology — Third-party verified #1 for accuracy
• 92 languages supported
• 40+ REST API endpoints — Full programmatic access for agent integration
• SOC-2 Type II & GDPR compliant
• Enterprise-grade security — Data never used for LLM training

In this agent, CustomGPT.ai serves as the knowledge base layer, providing semantic search, ranked context retrieval, and citation-backed answers from proprietary enterprise data.

"Business AI for trusted answers — no hallucinations, no guessing."

---

Team

Leadership

	Role
Alden Do Rosario	CEO, CustomGPT.ai

Project & Product Management

	Role & Contributions
Felipe Pires	Technical Product Manager & Project Manager — hiring and team building; task breakdown, distribution, and dispatch; evaluation strategy & methodology; regression/canary validation harness design; competitive benchmark research; multi-model integration and agent architecture direction

Developers

Developer	Focus
Aleksa Stojanović	Research, methodology, and team coordination — identified tooling and infrastructure gaps; built verification pipelines and analysis tooling; supported L3 edge-case tooling directions (Street View, multimedia, Wikipedia API); helped focus team effort on high-leverage work; led the final merge to main
Ramzi Mo	YouTube & media MCP tool suite (metadata, captions, song ID, frame/audio extraction) and orchestrator routing for video/audio; reasoning-rule additions (percent-change, formula delegation, count-with-exclusion, ranked extraction, temporal anchoring, multi-source/batch); RSV video-counting pattern; browser/PDF/figure error recovery; CRITIC verification rules (anti-sycophancy, FORMAT A/B, shortcut detection)
Hussein Younes	Vision pipeline — self-consistency voting in describe_image, the image_question narrow-question wrapper, describe-image caps / anti-fixation, and Gemini 3.1 vision routing for spatial questions; grounded-CRITIC verification; find_paper + fetch_via_wayback research tools; tool-delegation robustness and error-recovery fixes; quadruple-verification audit logging; and the agent test suite (voting, experience memory, infra-retry, subagents)
Arnav Gupta	Multi-provider orchestrator (Claude Opus 4.6/4.7 + GPT-5.5, per-run thinking level); vision-first computer use (headed Chromium + cu_* suite for Street View, 3D viewers, maps); Plan → Act → Verify loop (PLANNER + VISUAL ANALYST subagents, Gemini 3.1 Pro secondary vision); evidence tooling & observability (source-traceable research tools, model-I/O logging, real-time progress bus)
Dennis Yavuz	Early engineering on the initial implementation — orchestrator scaffolding, specialist subagents, MCP tool layer, and E2B sandbox integration

---

Acknowledgements

• Anthropic — Claude Opus 4.6 / 4.7, Claude Sonnet 4.6, and the Claude Agent SDK
• OpenAI — GPT-5.5 orchestrator provider
• Google — Gemini 3.1 Pro vision provider
• CustomGPT.ai — Enterprise RAG platform and knowledge base API
• E2B — Secure sandboxed code execution
• Playwright — Headed Chromium browser automation
• GAIA Benchmark — General AI Assistants evaluation framework

---

Citation

@misc{customgpt-agent-2026,
  title={CustomGPT Agent: Enterprise AI Agent with RAG-Augmented Multi-Agent Architecture},
  author={CustomGPT.ai Research},
  year={2026},
  url={https://github.com/adorosario/customgpt-agent}
}

---

_{Built with Claude Agent SDK and CustomGPT.ai}