Skip to content

Latest commit

 

History

History
59 lines (52 loc) · 5.51 KB

File metadata and controls

59 lines (52 loc) · 5.51 KB

Planner-free, Log-Mediated QA Architecture

Collaborative specialist agents coordinate through an append-only log to answer user questions while a verifier safeguards the final response.

┌───────────────┐      Read       ┌───────────────────────────────────────┐
│   User Query  │ ──────────────► │            Shared Append Log          │
└───────────────┘                 │  • Agent, Type, Content, Metadata     │
                                  │  • Read/Write history                │
                                  └───────────────────────────────────────┘
                                          ▲                     ▲
                                          │ Append              │ Read & Flag
┌─────────────────────┐        ┌──────────┴──────────┐          │
│ Specialist Agents   │        │  Controller Agent   │          │
│ ┌───────────────┐   │        │  • Turn-taking      │          │
│ │ Table Agent   │   │        │  • Simple scheduler │          │
│ ├───────────────┤   │        │  • Not a planner    │          │
│ │ Context Agent │   │        └──────────┬──────────┘          │
│ ├───────────────┤   │                   │                     │
│ │ Visual Agent  │   │        Append     │                     │
│ └───────────────┘   │                   │                     │
└─────────────────────┘                   ▼                     │
                                  ┌─────────────────────┐      │
                                  │ Summarizing Agent   │      │
                                  │  • Synthesizes      │      │
                                  │    candidate answer │      │
                                  └──────────┬──────────┘      │
                                             │ Answer          │
                                             ▼                 │
                                  ┌─────────────────────┐      │
                                  │ Verification Agent  │ ◄────┘
                                  │  • Cross-checks     │
                                  │    final response   │
                                  │  • Needs-Answer flag│
                                  └──────────┬──────────┘
                                             │ Verified Answer
                                             ▼
                                        ┌───────────────┐
                                        │     User      │
                                        └───────────────┘

Component Roles

  • Specialist Agents (Table, Context, Visual) extract and reason over domain-specific evidence. Each agent appends observations, intermediate reasoning, or lookup results to the shared log.
  • Controller Agent orchestrates the conversation by selecting which specialist should act next. It does not plan full solutions; instead it enforces orderly turn-taking and ensures the log progresses.
  • Shared Append Log is the central communication medium. Every agent reads from this append-only structure to understand context, then writes new entries capturing actions, findings, and requests for follow-up work.
  • Summarizing Agent monitors the log for a "need answer" flag. When triggered, it synthesizes a comprehensive answer using the accumulated evidence and appends the draft to the log.
  • Verification Agent reads the summarized answer, cross-checks against log evidence, and either confirms the final response or requests further investigation by writing a verification note back to the log.

Interaction Flow

  1. User submits a question. The controller logs the request and selects the appropriate specialist to begin evidence gathering.
  2. Specialists iterate via the log. Each specialist reads prior entries, performs retrieval or reasoning, then appends their findings. The controller assigns turns until sufficient evidence exists.
  3. Summarizer produces an answer. Once the log indicates a response is needed, the summarizing agent compiles the reasoning and proposes an answer in the log.
  4. Verifier cross-checks. The verification agent validates the proposed answer against log contents. If discrepancies exist, it flags the issue and may prompt additional specialist turns.
  5. User receives the verified answer. When the verification agent approves, the final answer is surfaced to the user along with any supporting rationale captured in the log.

This log-mediated collaboration enables robust multi-step reasoning without relying on a monolithic planner, while maintaining transparency and accountability for every decision taken by the agents.