Leashterm

Leashterm is a compact programming language designed specifically for AI agents, emphasizing security and bounded execution. Unlike traditional programming languages that allow unrestricted actions followed by sandboxing, Leashterm integrates essential constraints directly into its syntax, ensuring safer operation from the outset.

Key Features

• Explicit Permissions: Permissions to read or write data must be declared at the start of the program using the needs directive. Any attempt to access undeclared resources is blocked, enhancing security.
• Bounded Retries: The language allows retrying failed attempts using retry N { ... }, where N is a fixed number from 1 to 10. This prevents infinite loops and uncontrolled retrying.
• Verification Statements: With verify a == b, the program automatically halts if conditions do not hold, ensuring that only valid operations are executed.
• Tamper-Resistant Audit Log: Every action performed by the program is logged in a secure, hash-chained audit log that can be checked for tampering after execution.
• Structured Error Handling: Errors are reported in a structured JSON format, detailing the error type, line number, message, and hints for troubleshooting, accommodating smooth debugging and program repair.
• Guaranteed Termination: Each program is designed to terminate, with options for fixed repetition (retry) and iteration over finite lists (for), preventing endless loops.
• Work Limitations: Total execution work is restricted by a preset step budget, which is adjustable via --max-steps. This safety mechanism prevents unintended computational escalation, especially important when using nested operations.

Conceptual Framework

Leashterm is structured less as a general-purpose language and more akin to an executable capability manifest, tightly coupling predefined permissions with computation. The driving design philosophy is to minimize capabilities to ensure that each program admits a clear pre-execution upper bound on both its functionality and the work it can perform. This allows for effective capability checks before execution, maintaining control over agent actions while providing useful computational ability.

Example Syntax

Here's a brief example demonstrating Leashterm syntax:

# comment
needs read("notes.txt")    # declare permission to read
needs write("out.txt")      # declare permission to write

let text = read("notes.txt")  # read from file
print(text)                     # output the content

verify len(text) == 10          # check if text length is 10

retry 3 {                       # retry the block up to 3 times if failed
  let t = read("maybe.txt")    # attempt to read another file
}

if trim(text) == "yes" {       # conditional execution
  print(concat("got: ", text))
} else {
  print("no")
}

for f in ["a.txt", "b.txt"] {  # loop through provided files
  print(read(f))
}

Within this framework, the language values safety and accountability. It enforces strict computational limits while enabling meaningful work, striving to balance power and security in AI systems.