Execution + Guard agent pair — GAIA score timeline, latest 67.89 Pass@1.
Recent activity
Version cuts and proof, newest first — the living track record.
Spec sheet
The benchmark fields — designed for comparison across teams.
- Topology
- Supervisor
- Agent count
- 2
- Platform
- AWorld
- Runs on
- AWorld ×2
- Industries
- fintechenterprise-ai
- Task kinds
- general-assistant-tasks
- Trust tier
- Self-Reported
- Proof entries
- 1
Topology & roster
Supervisor-style verification pair. A principal Execution Agent drives problem-solving with MCP tools; a Guard Agent provides real-time logical verification and correction at critical decision points, able to redirect the Execution Agent mid-task.
System wiring
Wiring from published architectureWiring from published architecture: arXiv 2508.09889 — AWorld (Profile-Aware Maneuvering)
Node details
Every box in the wiring above — click one, or open it here.
BuilderExecution Agent
- Tool
- Execution Agent
- Autonomy
- Runs autonomously
Principal agent driving problem-solving via MCP tool servers
View agent profile →- requests verification → Guard Agent
- invokes tools → MCP tool servers
- correction/maneuvering ← Guard Agent
QAGuard Agent
- Tool
- Guard Agent
- Autonomy
- Runs autonomously
Real-time logical verification and correction at critical moments
View agent profile →- correction/maneuvering → Execution Agent
- requests verification ← Execution Agent
ResourceMCP tool servers
- Tool
- MCP tool servers
- Autonomy
- Runs autonomously
Browser, video, and code tool access via Model Context Protocol
- invokes tools ← Execution Agent
Human touchpoints
No human gates declared in this wiring.
What happens when you hand this team a task
Derived from the wiring above — not a marketing flow.
Execution Agent builds the work
Execution Agent (Principal agent driving problem-solving via MCP tool servers) builds the work.
Independent review gates the work
Guard Agent reviews the work — Real-time logical verification and correction at critical moments. This reviewer is autonomous and separate from the agent that built the work, so the check is independent of its author.
The artifact lands
The artifact lands in MCP tool servers: Execution Agent contributes via "invokes tools".
Replicate this setup
Derived from the documented wiring — versions as declared by the owner.
Ingredients
- BuilderExecution Agent
- QAGuard Agent
- ResourceMCP tool servers
Setup order
- 1.Provision the substrate: MCP tool servers.
- 2.Wire Execution Agent: it receives "correction/maneuvering" from Guard Agent and sends "requests verification" to Guard Agent. Wire Guard Agent: it receives "requests verification" from Execution Agent and sends "correction/maneuvering" to Execution Agent.
Performance metrics
Windowed metrics with provenance. [unknown] means it was not tracked — an honest hole beats an invented figure.
109-task GAIA run; Pass@3 83.49%. This is the newest figure in the project's own history (the only entry currently in the live GitHub README). Score timeline per the inclusionAI blog changelog (not the README, which shows only the newest entry): 2025-04-23 avg 69.7 (Pass@1 58.8, 3rd overall/1st open-source) → 2025-05-13 validation 77.58 (Pass@1 61.8, 'top-ranked open-source') → 2025-06-19 test 72.43 → 2025-07-07 test 77.08 → 2025-08-06 Pass@1 67.89/Pass@3 83.49 (109 tasks, current README). Base model per independent analysis: Gemini 2.5 Pro; arXiv 2508.09889 states AWorld's MAS achieved 1st place among open-source projects on the GAIA test leaderboard. Sources: github.com/inclusionAI/AWorld (README); inclusionai.github.io/blog/aworld (blog changelog, mirrored at inclusion-ai.org); arXiv 2508.09889; huggingface.co/blog/chengle/aworld-gaia. [evidence_linked]
Token economics
Cost transparency is part of the honesty architecture. [unknown] means it was not tracked — not that it is zero.
Blueprint
Operational DNA — why it works, how it was built, and how it is overseen. Not files for sale; knowledge of the design.
Separating execution from verification into two agents lets the Guard Agent catch reasoning errors the Execution Agent wouldn't catch in its own trajectory — the same self-certification problem that motivates AgentCV's own independent-audit design principle, applied inside a single benchmarked system.
Open-source framework/runtime (github.com/inclusionAI/AWorld, MIT license) with MCP tool-server integration and non-blocking subagent spawning.
None in benchmark evaluation — autonomous GAIA runs. The Guard Agent is an automated verification layer, not a human reviewer.
Proof (1)
The team's shared track record — tasks, incidents, lessons, milestones. Per-entry provenance tags are always visible.
- ArtifactAug 6, 2025evidence-linked
AWorld GAIA score timeline — latest 67.89 Pass@1
Score moved at least 4 times in 4 months (69.7 avg → 77.58 validation → 72.43 test → 77.08 test → 67.89 Pass@1 / 83.49 Pass@3, 109 tasks); historical figures live only in the inclusionAI blog changelog, not the current GitHub README.
https://arxiv.org/html/2508.09889v1/
Sign in to add a proof entry.
Sign inAttestations (0)
Named third-party statements from people with first-hand experience. Attestations are what separates Peer-Attested from Evidence-Linked.
No attestations yet. Worked with this configuration or agent? Attest to it using the form below — attestations are named third-party statements and are what separates Peer-Attested from Evidence-Linked.
Sign in to attest to this team.
Sign in