Client Background
AspiraTech is a Tokyo-based technology and consulting company supporting ambitious, technology-led organizations through business development, strategy, and cross-border market expansion , paired with the planning and operation of digital solutions using cutting-edge technology.
What stands out about AspiraTech is the combination of consulting-grade rigor and product-oriented execution. Their vision , "bringing ambitious ventures to new horizons with the power of technology (テクノロジーの力で、大志ある事業を新しい地平へ)" , is reflected in a portfolio that spans advanced AI initiatives and practical, value-driving services.
Challenge
Recruiting and headhunting workflows include a large amount of browser-based, repetitive work: navigating multiple web tools, extracting structured information, cross-checking records, and performing standardized updates. These tasks are often high-frequency and exception-heavy , making them poor fits for brittle, script-only automation.
AspiraTech wanted an approach that could:
- Operate on real web UIs reliably (dynamic pages, changing layouts, multi-step flows)
- Run in a securely isolated environment, given that recruitment operations can involve sensitive data
- Produce auditable outputs and execution traces suitable for business operations
- Scale from a pilot workflow to multiple similar workflows without redesigning the platform each time
Yodo Labs' Solution
Yodo Labs designed and delivered a cloud-sandboxed AI browser: an isolated "virtual browsing workspace" where an AI agent can perform end-to-end browser tasks under strict controls, producing structured outputs and run traces.
The system combined three key components:
An AI agent orchestration layer sits above the browser control plane , translating high-level goals into executable steps, selecting tools, validating page state, and recovering from common UI edge cases (timing, redirects, modal dialogs) without hard-coding every path.
A production-grade browser automation layer (Playwright)
We used Playwright as the deterministic control plane for the browser , enabling robust interaction with modern web apps, consistent session handling, and automation primitives that behave predictably across environments. [1]
- A secure cloud sandbox runtime (E2B) to isolate execution
To enforce isolation and provide an agent-safe execution surface, we ran the browser and agent runtime inside cloud sandboxes using E2B, which is designed for running untrusted or AI-generated workloads in secure, isolated sandboxes. E2B describes itself as open-source infrastructure for running AI-generated code in secure isolated sandboxes in the cloud, controllable via SDKs. [2]
This "AI browser in a sandbox" architecture gives a practical answer to a common enterprise requirement: let the agent work freely, but only inside a controlled blast radius.
How It Works (High-Level)
1) Workflow definition and guardrails
AspiraTech and Yodo Labs defined target tasks as goal-oriented workflows (rather than hard-coded click scripts), with clear constraints on what the agent is allowed to do.
2) Ephemeral sandbox provisioning (E2B)
In this implementation, each run was provisioned in a fresh sandbox environment. This creates strong separation across runs and reduces cross-session risk. While E2B supports persistent sandboxes, we chose an ephemeral-per-run model to maximize isolation for this use case. [2]
3) Agent-driven browsing (Playwright + tool use)
Inside the sandbox, the agent controls a real browser through Playwright. Playwright provides the reliable browser control primitives; the agent provides adaptive decision-making for handling UI variability and edge cases. [1] The agent maintains an internal task plan, checks completion criteria at each step, and escalates to human review for actions that cross defined risk thresholds.
4) Structured outputs + full run trace
Each run produces structured results intended for operational use (standardized fields, summaries, recommended next actions) along with an execution trace for review and continuous improvement.
Why Sandbox + AI Browser (Instead of Traditional Automation)
Traditional browser automation can become brittle when UI changes or exception paths multiply. Recruiting operations are particularly prone to these edge cases. The sandboxed AI browser pattern solves two problems simultaneously:
- Reliability: agent reasoning handles variability; Playwright ensures controlled execution. [1]
- Security & governance: the sandbox isolates the execution environment and supports safer rollout of agent capabilities. This is conceptually similar to remote browser isolation approaches [3], adapted here for server-side AI agent execution rather than end-user device protection.
Results
This project established a scalable foundation for operational automation in recruiting contexts:
- Reduced operational load by shifting repetitive browsing and data handling to an agent-driven workflow
- Improved consistency through standardized outputs and structured summaries
- Safer execution via sandbox isolation and controllable session lifecycles
- Enabled faster iteration on workflows, since new tasks could be defined by adjusting guardrails, prompts, and output schemas rather than rebuilding the platform layer
This architecture created a reusable foundation for onboarding adjacent workflows incrementally , while keeping the same security and operational model.
References
- Microsoft Playwright , Official documentation. playwright.dev
- E2B , Official documentation: Sandbox lifecycle & persistence. e2b.dev/docs/sandbox/persistence
- Cloudflare , What is browser isolation? cloudflare.com