temporal_agent_harness - the temporal_ai_agent but with more enterprise features
After Replay '26 and all the great customer conversations, we realized we should have a fuller-featured and modernized for 2026.
Customers are taking it and adding these features, we should just add them to the agent so it's ready to go. They are building agent platforms on it, so configurability and reuse is key.
Describe the solution you'd like
Ethan:
Users want us to provide agentic patterns and best practices.
We are increasingly asked to have an opinion about not just how to operate Temporal, but also how to achieve task success with agents.
A common theme: enterprise agent deployments are running into sandboxing, permissions, and auditability problems.
We should ship reference architectures for common patterns.
Showing developers how to build on Temporal will help convince them they should.
For example, there are a set of predictable sharp edges that come up when building long-running, interactive agents. We should have compelling, polished examples for these: long context window storage, sticky Activity execution, etc.
Features:
Security
- Guardrails pre and post LLM response
- user-based auth0 token passthrough and refresh - oauth based mcp access to secure data
- PII filtering hooks from user input and tools output
- use small models (LLMs as judge) to validate/guardrail requests and responses
- session analysis (long sessions may indicate system abuse)
(Guardrails on output streaming is tricky, need to figure that out, maybe something deterministic)
Feature/improvements
- better MCP integration
- dynamic tool loading
- system prompts and context as input/config
- observation (Temporal? Langfuse?)
- conversations export for analysis
- conversation storage outside of workflow, e.g. database
- interruptible chat/tools
- multi media output/UI
- large payload support for conversations
- streaming responses? At least a thinking message and token count
- improved UI
- approval memory framework
- evals
- model fallback
Additional context
temporal_agent_harness - the temporal_ai_agent but with more enterprise features
After Replay '26 and all the great customer conversations, we realized we should have a fuller-featured and modernized for 2026.
Customers are taking it and adding these features, we should just add them to the agent so it's ready to go. They are building agent platforms on it, so configurability and reuse is key.
Describe the solution you'd like
Features:
Security
(Guardrails on output streaming is tricky, need to figure that out, maybe something deterministic)
Feature/improvements
Additional context