Skip to content

agent-roadmap.md

Latest commit

 

History

History
78 lines (50 loc) · 2.48 KB

agent-roadmap.md

File metadata and controls

78 lines (50 loc) · 2.48 KB

Agent Feature Roadmap

Docs Home | Agent-Native | Benchmarks | Framework Comparison | Ecosystem

This roadmap tracks high-value work to strengthen Qirrel as an agent-compatible runtime while preserving lightweight deterministic parsing.

Roadmap Principles

  • Keep deterministic parsing as a stable core.
  • Add agent/runtime capabilities as optional layers.
  • Prefer explicit contracts (schemas, errors, protocol behavior).
  • Avoid hidden coupling between direct API mode and agent mode.

Priority Backlog

1. MCP Resources and Prompts

Why: improves interoperability and discoverability in MCP-native orchestrators.

Candidate scope:

  • resources/list, resources/read
  • prompts/list, prompts/get
  • reusable extraction profile resources

Done when: tools + resources + prompts can be consumed by a generic MCP client without custom glue.

2. Strict Tool I/O Contracts

Why: tool invocation failures should be deterministic and machine-actionable.

Candidate scope:

  • strict schema validation for tools/call arguments
  • optional output-schema enforcement
  • standardized validation error payloads

Done when: invalid inputs consistently return structured, predictable error responses.

3. Memory Adapters

Why: many agents need short-term and long-term memory surfaces.

Candidate scope:

  • in-memory session memory
  • SQLite adapter
  • key-value interface for external stores

Done when: memory adapters are pluggable without changing parse pipeline internals.

4. Human-in-the-Loop Controls

Why: operations teams need explicit approval controls for sensitive flows.

Candidate scope:

  • tool-level approval metadata
  • pause/resume workflow hooks
  • policy hooks (allow/deny by tool and argument shape)

Done when: approvals can be enforced by policy without forking core runtime.

5. Tracing and Evaluations

Why: production agent stacks need observability and reproducible quality checks.

Candidate scope:

  • event-to-trace export (JSONL/OTel-friendly)
  • eval harness for tool success rate/latency/schema violations

Done when: teams can compare versions with repeatable metrics and trace artifacts.

Suggested Sequencing

  1. strict tool contracts
  2. resources/prompts
  3. tracing/evals
  4. memory adapters
  5. human approval controls

This order front-loads interoperability + reliability before higher-level workflow features.