Karthikvenugopal

AgentScope — Development skill for Claude Code

Development community

Research harness for evaluating Codex and Claude Code across single, staged, and parallel agent architectures with isolated execution, independent verification, hierarchical telemetry, Harbor interope.

How to install AgentScope

This entry records only its repository, not the path inside it, so there is no exact command to give. Open Karthikvenugopal/AgentScope and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What AgentScope does

Alternatives in Development

Maestro Next — Unified entry for all development intents — classify intent, assess complexity, route to the correct execution 530 ★
Context Engineering Collection — Use when building, optimizing, or debugging agent systems that require effective context management — a compre 176 ★
Music Genre Finder — 🎵 Intelligent music genre search with 5947 genres from RateYourMusic - Claude Code skill for quick lookup, sm 158 ★

README

AgentScope

**Release 1.0.0 · completed research-engineering system**

AgentScope is a research-oriented coding-agent harness for measuring how execution architecture changes software-engineering correctness and provider workload. It runs Codex CLI or Claude Code against reproducible repositories, records hierarchical agent and tool traces, and applies an independent verifier after the candidate patch is frozen.

It exists to study questions that a model-only benchmark cannot answer: What does a planning or review stage cost? Do parallel candidates overlap enough to reduce wall time? How many provider invocations and tool calls does an architecture create? Does added orchestration change official task success when provider, task, and environment are held constant?

AgentScope measures harness-visible provider processes, tokens where available, tools, patches, latency, and official verification. It does **not** measure GPU utilization, true TTFT/ITL, or inference-server efficiency.

[![Completed controlled experiment: clustered strategy estimates and task matrix](docs/images/phase9-strategy-summary.png)](docs/images/phase9-strategy-summary.png)

Read the compact systems report: **[AgentScope: Evaluating the Correctness and Inference-Workload Trade-offs of Coding-Agent Harness Architectures](docs/research-report.md)**.

What it can do

Run Codex CLI, Claude Code, or a deterministic mock provider behind capability-aware adapters.
Execute single-agent, planner → implementer → reviewer, and bounded parallel-implementer strategies.
Give every candidate an isolated Docker workspace with typed, confined tools.
Keep hidden tests, reference solutions, sibling workspaces, Docker authority, and database credentials away from agents.
Select a parallel candidate before running the only official verifier.
Import current Harbor-style tasks without delegating AgentScope orchestration to Harbor.
Persist runs, sub-executions, traces, metrics, candidate selecti