How It Works
HumanLayer originally provided an SDK designed to integrate human approval workflows into AI agent systems. This SDK allowed agents to pause execution of high-stakes tool calls and request explicit human input or validation before proceeding. The approval process could be routed through multiple channels, including Slack, email, or a dedicated web UI, ensuring flexibility in how oversight was conducted. Once a human decision was made, the agent would resume its workflow based on the provided input.
The SDK functioned as a runtime defense mechanism, ensuring critical actions were not automated without human intervention.
While HumanLayer has since shifted its primary focus to coding-agent tooling under the CodeLayer brand, the original approval SDK remains a foundational reference for implementing human-in-the-loop oversight in AI systems. The core concept—pausing agent execution for human validation—serves as a template for other developers building similar safeguards into their own AI workflows.
The SDK was built using TypeScript and released under the Apache 2.0 license, ensuring permissive use and integration with a wide range of projects. Its architecture was designed to be lightweight yet extensible, accommodating different approval channels and decision-making processes.
Use Cases
The approval SDK was primarily intended for AI agents operating in environments where automated decisions carried significant risk. For example, an AI agent managing patient records might require human approval before modifying critical data fields, ensuring adherence to privacy laws like HIPAA.
In enterprise settings, the SDK could be deployed to enforce governance policies, such as requiring managerial approval for budget-related tool calls or legal review for contract-generating agents. This level of control was particularly useful in large organizations where decentralized AI tools might otherwise operate without sufficient oversight.
By routing approval requests through familiar channels like Slack or email, teams could maintain productivity while ensuring critical decisions were not left to automation alone.
Key Features & Capabilities
The approval SDK offered several distinct features to facilitate human oversight in AI workflows.
- Multi-channel Approval Routing: Agents could request human input via Slack, email, or a custom web interface, depending on the operational context and team preferences. This adaptability ensured that approval processes could be integrated seamlessly into existing communication tools.
- Pause-and-Resume Execution: The SDK allowed agents to halt their current task, wait for a human decision, and then resume execution based on the provided feedback.
- Modular Integration: The SDK was designed to be embedded into existing AI agent frameworks with minimal disruption. Developers could incorporate approval workflows into their pipelines without requiring a complete redesign of their systems.
- Apache 2.0 License: The permissive licensing model encouraged adoption across open-source and proprietary projects, fostering broader use and collaboration within the AI community.
Beyond the approval SDK, HumanLayer’s current focus on CodeLayer introduces additional capabilities tailored to AI-driven software development. While the original SDK remains relevant for runtime defense, CodeLayer extends these principles into collaborative coding environments.
- Multiplayer Coding Agent IDE: CodeLayer functions as an integrated development environment (IDE) that supports collaborative agent sessions, allowing teams to work together in real-time on coding tasks. This includes shared workspaces where agents and humans can co-edit code, review design documents, and manage artifacts.
- Task and Artifact Management: The platform groups agent sessions, code artifacts, and worktrees into structured tasks, enabling teams to track progress, provide feedback, and maintain alignment throughout the development process. Each task can include design documents, code diffs, and acceptance criteria, ensuring transparency and accountability.
- Six-Phase Workflow for Code Development: CodeLayer implements a structured approach to turning high-level intent into shipped code, with checkpoints at each phase. These phases include:
- Generating clarifying questions before writing code to ensure alignment on requirements.
- Mapping the codebase and dependencies to understand the context.
- Drafting design documents that teams can review and comment on in real-time.
- Breaking designs into verifiable implementation steps with detailed plans.
- Executing code with continuous feedback to catch issues early.
- Ensuring no critical thinking is outsourced to automation, with humans remaining in control of quality and architecture.
- Integration with Existing AI Models: CodeLayer supports a variety of AI models, including Claude Code, Codex, Copilot, and Fireworks, allowing users to bring their own AI subscriptions (BYOK). This flexibility ensures compatibility with existing toolchains and reduces dependency on proprietary token billing.
- Collaborative Design Reviews: Unlike traditional tools like Notion or Google Docs, CodeLayer ties design documents directly to the implementation process. Comments and decisions made during design reviews are fed back into the agent workflow, ensuring that feedback is actionable and integrated into the final codebase.
- Scalable Agent Orchestration: The platform supports local and cloud-based agent sessions, enabling teams to run parallel sessions on workstations or scale to cloud environments for long-running tasks. All sessions sync through the HumanLayer API, providing a consistent experience across devices and collaboration tools.
- Security and Compliance: CodeLayer is SOC 2 Type II compliant, with additional security measures such as ZDR (Zero-Day Response) and DPA (Data Processing Addendum) agreements available for enterprise tiers.
- Open-Source Framework: While CodeLayer itself is not fully open-source, components of its Research, Plan, Implement (RPI) framework are available under open-source licenses.
The transition from the approval SDK to CodeLayer reflects a broader trend in AI tooling: moving from standalone oversight mechanisms to integrated, collaborative environments where humans and agents work in tandem. The approval SDK’s principles of human validation remain embedded in CodeLayer’s design, ensuring that even in automated coding workflows, critical decisions retain human oversight.