Does this work with Anthropics new Harness for long running agent strategy?

[ronval]

Hey very cool project/repo you made.

I was wondering if this would work with Anthropics new long running agent strategy. They made a blog post about it here
https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents

and this is the autonomous coding project example

https://github.com/anthropics/claude-quickstarts/tree/main/autonomous-coding

Would this SDK work with that or does the autonomous-coding project need its own ruby version as well?

the autonomous-coding project depends on the Agent SDK so not sure if this ruby version is a complete port?

Any insights you have would be great to know.

Cheers :)

[ya-luotao]

Hey, thanks for the question!

Short answer: Yes, this SDK works with the long-running agent pattern.

I dug into the https://github.com/anthropics/claude-quickstarts/tree/main/autonomous-coding and here's the thing — most of what makes it "long-running" isn't actually in the SDK. It's application code:

• feature_list.json for tracking progress across sessions
• A bash command allowlist for security
• Session management logic that loops and resumes

The SDK just provides the building blocks: Client, query(), receive_response(), hooks, etc. This Ruby SDK has all of those — same API as the Python one.

So you wouldn't need a separate Ruby SDK. You'd just write the application layer in Ruby using this SDK. Here's a rough sketch:

require 'claude_agent_sdk'

# Run sessions in a loop, persisting progress to a file
loop do
  options = ClaudeAgentSDK::ClaudeAgentOptions.new(
    cwd: './my_project',
    allowed_tools: %w[Read Write Bash Glob Grep],
    hooks: {
      'PreToolUse' => [
        ClaudeAgentSDK::HookMatcher.new(
          matcher: 'Bash',
          hooks: [->(input, _, _) {
            # Block commands not in allowlist
            cmd = input.input[:command].to_s.split.first
            %w[ls cat grep git npm node].include?(cmd) ? nil :
              ClaudeAgentSDK::PreToolUseHookOutput.new(decision: 'block')
          }]
        )
      ]
    }
  )

  ClaudeAgentSDK.query(prompt: load_prompt, options: options) do |msg|
    # Process messages, update feature_list.json as features complete
  end

  break if all_features_done?
  sleep 3  # Auto-continue
end

The pattern is the same as the Python demo — just Ruby syntax.