Skip to main content

Roadmap

This roadmap captures recommendations from research on making libraries easier for agents to use. Items are organized by their complexity-to-value tradeoff: high-value, low-complexity items first.

High Value, Low Complexity​

Verbose Error Messages with Recovery Instructions​

Status: Partially implemented

Every error should include a plain-English explanation and at least one suggested fix. Research shows that chatty text errors drive faster agent recovery than terse codes alone.

"While JSON structure is fine for machines, a plain-English 'message' field often helps an LLM more than a perfectly terse schema." — Stytch

Current state: DIAL errors include codes and messages. Some errors include recovery suggestions.

Gap: Not all error paths include actionable recovery instructions. Need systematic audit.

OpenAPI Schema for HTTP Transport​

Status: Not implemented

Publish an OpenAPI 3.1 schema for the HTTP transport endpoints. Agents using structured tool calling can directly ingest the schema.

"When creating action groups for agents, you must define the parameters using an OpenAPI schema in JSON or YAML format." — AWS Bedrock Docs

Value: Enables automatic tool generation for frameworks like Amazon Bedrock, LangChain, and others that consume OpenAPI.

Document All Error Codes Exhaustively​

Status: Not implemented

Create a reference page listing every error code, what triggers it, and how to fix it. Agents need this to fail forward instead of stalling.

"Every error an endpoint can throw should be documented, listing the code, triggering conditions, and at least one fix." — Stytch

High Value, Medium Complexity​

Bulk Operations for Sessions​

Status: Not implemented

Add batch endpoints for common operations: create multiple sessions, solicit proposals from all proposers in one call.

"Wrapping bulk APIs over single-entity APIs enables LLM/agent efficiency, preserves backward compatibility, and keeps systems secure and extensible." — Stytch

Current state: Each operation is a single call. Running 10 proposals requires 10 API calls.

Gap: For multi-agent scenarios, this creates unnecessary round-trips.

Rich Tool Descriptors for MCP​

Status: Partially implemented

Enhance MCP tool schemas with detailed parameter descriptions, expected outcomes, and potential side effects.

"The March 2025 MCP specification update introduced richer tool descriptors, allowing servers to provide more detailed metadata about tool parameters, expected outcomes, and potential side effects." — MCP Best Practices

Current state: MCP tools have basic descriptions.

Gap: Missing detailed parameter documentation, outcome descriptions, and side effect warnings.

llms.txt MCP Integration​

Status: Not implemented

Create an MCP server that serves llms.txt content on demand, allowing agents to fetch relevant documentation sections as needed.

"LLMTEXT, an open source toolkit, helps developers create, validate, and use llms.txt files—making any website instantly accessible to AI agents through standardized markdown documentation and MCP servers." — Parallel.ai

Value: Agents can load documentation into context window on demand rather than pre-loading everything.

Medium Value, Low Complexity​

Examples in Every Error Message​

Status: Not implemented

Include a valid example alongside each error message showing the correct format.

Current state: Errors describe what's wrong.

Gap: Errors don't show what right looks like.

Tool Count Optimization​

Status: Implemented (good state)

Keep the number of tools small and focused. DIAL exposes 12 tools, each with a single purpose.

"Fewer, well-designed tools often outperform many granular ones, especially for agents with small context windows or tight latency budgets." — MCP Best Practices

Current state: DIAL already follows this principle. No action needed.

Machine Definition Validation Endpoint​

Status: Not implemented

Add a validateMachine tool/endpoint that checks a machine definition for errors before creating a session.

Value: Fail fast before execution. Agents can iterate on definitions without wasting cycles.

Medium Value, Medium Complexity​

Reflection/Self-Correction Patterns Documentation​

Status: Not implemented

Document how to build self-correcting agents using DIAL, where a specialist evaluates its own output and refines it.

"The Evaluator-Optimizer pattern uses a self-correction loop where a second LLM step acts as a reflector or evaluator, critiquing the initial output." — AIMultiple Research

Value: Guidance for building more robust specialist implementations.

Streaming Responses for Long Operations​

Status: Not implemented

Expose streaming for operations that may take time, such as consensus evaluation with many specialists.

Current state: No SSE or streaming transport is implemented yet. The HTTP server uses JSON-RPC over POST.

Gap: Long-running operations should emit progress events via SSE or similar streaming transport.

Medium Value, High Complexity​

Agent-to-Agent (A2A) Protocol Support​

Status: Not implemented

Implement the emerging A2A protocol for multi-agent communication, allowing DIAL specialists to communicate directly.

"In multi-agent systems, agents communicate using Agent-to-Agent (A2A) protocols that define the flow of information between them." — AIMultiple Research

Complexity: Requires protocol implementation, security model, message routing.

Value: Native support for complex multi-agent workflows.

OAuth 2.1 Authorization​

Status: Not implemented

Replace simple token auth with OAuth 2.1 for production deployments.

"The March 2025 update to the MCP specification formally recommends OAuth 2.1 as the primary mechanism for authorization." — MCP Best Practices

Complexity: OAuth flows, token management, refresh handling.

Value: Enterprise-grade security, standardized auth for MCP ecosystem.

Code Execution for Tool Composition​

Status: Not implemented

Allow agents to write code that composes DIAL tools rather than calling each tool individually.

"Code execution with MCP enables agents to use context more efficiently by loading tools on demand, filtering data before it reaches the model, and executing complex logic in a single step." — Anthropic Engineering

Complexity: Sandboxed execution environment, security model.

Value: Dramatically reduces context usage for complex operations.

Low Value, Low Complexity​

Semantic Versioning in Responses​

Status: Not implemented

Include the DIAL version in API responses so agents know what spec they're working with.

Current state: Version is in package.json but not in API responses.

Gap: Agents must read package.json to know the version.

Health Check Endpoint​

Status: Not implemented

Add a /health endpoint for the HTTP transport that returns server status.

Value: Standard infrastructure pattern. Low effort.

Contributing​

If you're implementing any of these items, please:

  1. Update this document to reflect the new status
  2. Add documentation following the spec change workflow
  3. Write changelogs as instructions to agents

Sources​

Research informing this roadmap: