AutoBot: A Voice-Driven Operating Harness for Persistent AI Workflows

AutoBot is a privacy-centric agentic framework designed to enhance the capabilities of frontier AI models, specifically focusing on the completion of intricate, long-term knowledge work. By integrating directly with native ChatGPT Project features—including synchronous voice, the OAI UI, and various control planes—it transforms a standard LLM into a robust, self-improving worker.

🚀 Performance & Benchmarks

AutoBot significantly outperforms existing baselines in complex, multi-application environments.

• OSWorld 2.0: In benchmarks testing long-form workflows across multiple apps, AutoBot achieved a $\approx 18.5\%$ increase in task completion over the OpenAI Sol Max baseline, even surpassing Anthropic’s Claude Opus 5 Max.
  • Binary Accuracy: $32.41\%$
  • Partial Accuracy: $64.28\%$
• AssistantBench: AutoBot currently holds the #1 spot on the official hidden-test leaderboard with an accuracy of $50.70\%$ across 181 distinct tasks.

AutoAssist $\rightarrow$ AutoBot (The project was rebranded two weeks after the initial benchmark screenshots).

🛠️ Technical Methodology

The core philosophy of AutoBot is that difficult workflows should serve as upgrades to the agent itself.

The Self-Improvement Cycle

AutoBot doesn't just execute tasks; it maintains its own infrastructure. It can independently repair its harness, validate those repairs, and implement them permanently.

Memory and State Management

To prevent the "forgetting" common in long chats, AutoBot utilizes a sophisticated memory architecture:

1. Hierarchical Disk Storage: Memory is persisted on disk rather than relying solely on context windows.
2. Task-Specific Retrieval: Only relevant context is pulled into the active working memory.
3. Nightly Consolidation: The system performs a scheduled merge of new knowledge and corrections every night.
4. Institutional Memory: Knowledge accumulates across different projects and separate conversations.

Reliability and Efficiency

• Resumability: Through the use of atomic checkpoints, persistent task graphs, and independent supervision, a new worker can step in and resume a task exactly where the previous one left off.
• Verification: Completion is not assumed; it is bound to specific requirements and verified via "destination evidence."
• Inference Optimization: By using compiled context and reusable proofs, AutoBot minimizes redundant computation, allocating the model's "reasoning budget" to high-stakes judgments.

🔒 Privacy and Security Framework

AutoBot implements a strict "Privacy by Relationship and Purpose" model.

Contextual Zoning

Data is partitioned into distinct zones to prevent leakage:

• Personal
• Family & Friends
• Work
• Opt-in Shared Context

The Validation Contract

To ensure integrity, AutoBot employs a strict separation of duties:

The Producer/Validator Split: The local runtime forbids a single entity from both producing a result and validating it. A separate validator must inspect the evidence—including the actual destination—whenever a task modifies something outside the immediate workspace.

Operational Safeguards

• Persistence: Objectives, stages, and recovery states are saved to disk so that interrupted sessions do not result in lost commitments.
• Write Verification: All recipient-visible writes occur via signed-in first-party apps using Computer Use.
• The Verification Loop: The workflow must verify the $\text{Account} \rightarrow \text{Destination} \rightarrow \text{Content}$ sequence before acting, then inspect the rendered result to ensure there are no duplicates or unwanted "AI-generated" signatures. If this "clean route" fails, the process stops immediately.
• Communication: Instead of raw archives, AutoBot stores compact, user-approved patterns.

📦 System Components

AutoBot is structured as a two-layer stack:

Layer 1: Native ChatGPT

• Desktop Application
• Local Projects & Goals
• Synchronous Voice
• Skills & Plugins
• Scheduled Work & Notifications
• Computer Use capabilities

Layer 2: The AutoBot Workspace

• Operating contract defined in AGENTS.md
• Privacy zone management
• Selective memory systems
• Local objective state machine
• 60-second liveness supervisor
• External-action policy
• First-time setup scripts

Note: Functionality is dependent on the user's specific ChatGPT plan, region, and permission limits.

📊 Comparative Analysis

The following table compares AutoBot (integrated with native ChatGPT) against OpenClaw and Hermes for a single Mac user managing professional and personal tasks.

Definitions:

• "Better": A more explicit default for the stated user need.
• "Doesn't do": Requirement not found in primary documentation (as of Sept 10, 2026).

Implementation Example: The Operating Contract

The logic governing these agents is often stored in a structured format like this:

# Example snippet from AGENTS.md
policy:
  external_write:
    require_verification: true
    validator_label: "independent_supervisor"
    check_rendered_output: true
    stop_on_attribution_failure: true