Project summary

AI agents are moving from chat interfaces into real work: running terminal commands, editing files, calling tools, and increasingly coordinating with other agents. The safety layer around that work is still immature. Many builders rely on raw terminals, ad hoc logs, and informal message passing, which makes it hard to catch risky behavior while it is happening or to control how agents communicate across ownership boundaries.

SJK Systems builds middle-layer infrastructure for that gap. ClevAgent is a supervised terminal for agent work: the agent works inside the terminal, risky or wasteful behavior is detected in the active session, and corrective guidance is delivered before a small mistake becomes a larger problem. ForAgent is telecom for AI agents: it gives each agent a private number so agents can call each other through a controlled app, with owner approval, logs, revocation, and rate limits around every call.

This grant request is not general company runway. It funds public-benefit artifacts other builders can use.

What are this project's goals? How will you achieve them?

Goal 1: Publish a practical taxonomy of risky AI coding-agent terminal behaviors: destructive commands, over-broad file access, secret exposure, prompt-injection-like behavior from tool output, task drift, and unsafe permission escalation.

Goal 2: Build a synthetic scenario set and evaluation harness for supervised terminal sessions. Scenarios use no real credentials, private repositories, or customer data.

Goal 3: Implement reference controls that show how a supervised terminal can detect and correct risky behavior while work is happening.

Goal 4: Publish an initial safety specification for controlled agent-to-agent calls, using ForAgent's telecom model: private numbers, explicit connection approval, call logs, revocation, rate limits, and human-readable audit trails.

Goal 5: Release implementation notes and documentation so small teams can adopt the work without enterprise security resources.

How will this funding be used?

At $5K: the terminal-risk taxonomy, synthetic scenario plan, and initial supervised-terminal control design. At $15K: the evaluation harness, initial detection and correction controls, and a draft agent-call safety specification. At $25K: the complete harness and documentation, model-assisted evaluations across the scenario set, adversarial tests for unsafe agent communication, and limited external security review.

Funds support engineering time, documentation, model-evaluation costs, and security-oriented review. All intended outputs are public artifacts.

Who is on your team?

Solo founder: Sean Kwon, founder and CEO of SJK Systems, Inc., a bootstrapped Michigan C-Corp incorporated in April 2026. His background is in finance and investment brokerage at Deutsche Bank and J.P. Morgan, where real work depended on controlled systems between intention and execution: trust boundaries, logs, and auditability. He holds an MBA from the University of Michigan Ross School of Business and built and launched the ClevAgent and ForAgent products.

What are the most likely causes and outcomes if this project fails?

The main risk is that the public artifacts are too abstract to be useful. To reduce it, the deliverables stay concrete: synthetic traces, a runnable harness, reference controls, and implementation notes. A second risk is that model-assisted detection is unreliable in ambiguous terminal sessions. The project will compare deterministic rules, model-assisted review, and hybrid approaches, and will document where model-based classification should not be trusted.

How much money have you raised in the last 12 months, and from where?

$0 in outside cash funding. SJK Systems is bootstrapped, with infrastructure credits from Microsoft for Startups and Datadog for Startups.