Skip to content

03-harness-graph-patterns.md

Latest commit

 

History

History
70 lines (50 loc) · 2.39 KB

03-harness-graph-patterns.md

File metadata and controls

70 lines (50 loc) · 2.39 KB
Raw

03 — Harness Graph Patterns

Four recurring harness topologies. Each is connected back to the theory, and each has a runnable lab in ../labs.

Pattern 1 — Single-Agent Wrapper

[Task] -> [Agent / LLM] -> [Output] -> [Verifier] -> [Accept?]

Theory. One path. Few observable internal boundaries. The bottleneck is often hidden inside the agent. Low trace resolution: when output is wrong, the trace cannot localize the cause.

Demo: 01-single-agent-wrapper.

Pattern 2 — Context-Aware Harness

[Task] -> [Context] -> [Agent / LLM] -> [Verifier] -> [Accept?]
             ^
             |
      [Repo / Docs / Memory]

Theory. Adds an inspectable context edge. We can measure context relevance, context loss, stale memory, and missing information. Common bottleneck: Context -> Agent.

Demo: 02-context-aware-harness.

Pattern 3 — Feedback Harness

[Task] -> [Context] -> [Agent] -> [Verifier] -> [Accept?]
                          ^            |
                          |            |
                          +---fail-----+

Theory. Adds a feedback edge. Error becomes signal. The loop can improve, oscillate, or saturate. Useful for debugging convergence.

Demo: 03-feedback-harness.

Pattern 4 — Parallel / Multi-Path Harness

                /-> [Path A] -\
[Task] -> [Shared Context]      -> [Compare / Verify] -> [Accept?]
                \-> [Path B] -/

Theory. Multiple paths carry candidate flow. Compare quality/cost/error and choose the path with the highest useful-flow score Φ. Can reveal bottlenecks and route around weak paths.

Demo: 04-parallel-multipath-harness.

What can be inferred before implementation

The graph exposes tradeoffs before any agent code is written:

Pattern Pros Cons / new failure mode
Single-agent wrapper simple, low coordination cost low observability, hidden bottlenecks
Context-aware harness better input shaping, measurable context quality wrong / stale context
Feedback harness enables improvement requires good error signals; can oscillate or saturate
Parallel harness explores alternatives, higher chance of accepted output higher comparison and coordination cost