DurableAgent missing feature parity with AI SDK Agent class

[kvnwolf]

Description

The DurableAgent class in the Workflow Development Kit is missing several important configuration options that are available in the AI SDK's ToolLoopAgent class. This makes it difficult to migrate existing agent implementations to durable workflows while maintaining the same level of control and customization.

Current Behavior

DurableAgent currently only supports three configuration options:

• model
• tools
• system

Expected Behavior

DurableAgent should support feature parity with ToolLoopAgent, or at least provide access to the most commonly used configuration options.

Missing Features

The following configuration options from ToolLoopAgent are currently not available in DurableAgent:

Critical Missing Features

1. providerOptions - Essential for configuring provider-specific settings (e.g., custom headers, API configurations)
2. maxRetries - For configuring retry behavior on failed requests
3. abortSignal - For cancellation support
4. experimental_telemetry - For observability and monitoring
5. onStepFinish - For step-level hooks and observability

Generation Control

6. maxOutputTokens - Token limit control
7. temperature - Response randomness control
8. topP - Nucleus sampling
9. topK - Top-k sampling
10. presencePenalty - Repetition penalties
11. frequencyPenalty - Frequency penalties
12. stopSequences - Custom stop sequences
13. seed - Deterministic generation

Tool Control

14. toolChoice - Control over tool selection behavior
15. stopWhen - Custom stop conditions for the agent loop
16. activeTools - Dynamic tool enabling/disabling
17. prepareStep - Step preparation hook
18. experimental_repairToolCall - Tool call repair functionality

Output Control

19. output - Structured output with schema validation

Use Case

I need several of these options in my durable agents, particularly:

• providerOptions - To configure custom provider settings (e.g., custom endpoints, headers)
• temperature and other generation params - To control response quality and determinism
• maxRetries - To handle transient failures gracefully
• onStepFinish - For monitoring and debugging agent behavior
• experimental_telemetry - For production observability

Code Comparison

What works with ToolLoopAgent (AI SDK):

import { ToolLoopAgent } from 'ai';

const agent = new ToolLoopAgent({
  model: openai('gpt-4o'),
  instructions: 'You are a helpful assistant.',
  tools: { myTool },
  providerOptions: {
    openai: {
      headers: { 'Custom-Header': 'value' }
    }
  },
  temperature: 0.7,
  maxRetries: 3,
  onStepFinish: ({ step }) => {
    console.log('Step finished:', step);
  },
  experimental_telemetry: {
    isEnabled: true,
    functionId: 'my-agent'
  }
});

What's currently not possible with DurableAgent:

import { DurableAgent } from '@workflow/ai/agent';

const agent = new DurableAgent({
  model: 'anthropic/claude-haiku-4.5',
  system: 'You are a helpful assistant.',
  tools: { myTool },
  // ❌ providerOptions: {...},  // Not available
  // ❌ temperature: 0.7,  // Not available
  // ❌ maxRetries: 3,  // Not available
  // ❌ onStepFinish: {...},  // Not available
  // ❌ experimental_telemetry: {...},  // Not available
  // ... and many more options
});

References

• DurableAgent Documentation
• ToolLoopAgent Reference

Workaround

Currently, there's no clear workaround to access these options when using DurableAgent.

Would it be possible to either:

1. Add these configuration options to DurableAgent, or
2. Provide guidance on how to achieve similar functionality with the current API?

Having at least the most critical options (providerOptions, temperature, maxRetries, onStepFinish, and experimental_telemetry) would significantly improve the developer experience when building production-grade durable agents.

Thank you!

[pranaygp]

Great call and yeah, the current implementation is a barebones version to just get streaming up and running!

Thanks so much for making this task. I'm working on getting closer to feature parity

[pranaygp]

adding this to the list to track work

Model name is serialised, can't wrap up tracing around the LanguageModel. Doesn't support stopWhen, experimental_context, prepareToolCall, prepareStep. Workflow doesn't support https locally.

source: https://x.com/gdaybrice/status/1989099115087818769?s=46

[gdaybrice]

FYI, if that helps anyone in the meantime,

With a bit of plumbing, you can support resumable streams with the latest ToolLoopAgent (tools are not steps, but it's fine in my current implementation).

https://x.com/gdaybrice/status/1996748681102741582

[pranaygp]

Posting this snippet here from @nicoalbanese with some more issues for posterity and visibility. Will need to kick off an agent soon to try and work through a lot of the issues. I did just ship client side tool support in #1329

---

My pain points with durableagent and workflow:
• no call options  (which are a fundamental part of my current implementation)
• no type helpers (InferAgentUIMessage)
• doesn't support client-side tools?
• prepareStep is at the call site and not in the agent definition?
• given serialisation requirement between workflow → steps requires a different way of building versus traditional toolloopagent -> requires rewriting tools to reconnect to sandbox in each one
• persisting quite painful → collect ui message → not typed → differences from existing onFinish in toUIMessageStream approach
• constant errors from using node in workflow due to missing "use steps" but incredibly hard to read and find exactly where
• aborting is non-trivial -> do you cancel a workflow? but then it just stops, and doesn't persist steps that were in progress
• initial step loading time (ttfb) for each run is now taking 5 seconds due to sandbox reconnection logic (skill issue that I need to work through) [mostly fixed, but was not happy path]

[pranaygp]

Just merged #1362 which brings us a lot closer to parity with AI sdk, and adds tests cases for the remaining gaps that we can more easily address now