agentsys

A modular runtime + orchestration system for AI coding agents — 24 plugins, 49 agents, 44 agent-skills — for Claude Code, OpenCode, Codex CLI, Cursor, and Kiro. The clearest statement of the agentic-coding-harness thesis: frontier models can write code; the open problem is “everything else” — task selection, branch management, review, CI/CD, deployment.

Architecture principles

• Separation of concerns — each agent has a single responsibility and a specific model assignment (cheaper models where they suffice).
• Tool-based efficiency — deterministic work (static analysis, regex, AST parsing) is done with tools, not LLM tokens.
• Certainty grading — findings tagged HIGH/MEDIUM/LOW confidence to route auto-fix vs. manual review (the same honesty/uncertainty discipline claude-opus-4-8 foregrounds).
• Phase gates — pipelines enforce ordered steps so agents can’t skip testing/review; the execution-time sibling of spec-driven-development.

Capabilities

Slash commands: /next-task (end-to-end workflow), /prepare-delivery (quality gates), /deslop (strip AI artifacts), /audit-project (multi-agent review). State persists across sessions so interrupted work resumes. Claimed benchmark: Sonnet + agentsys beats raw Opus on cost-effectiveness by leaning on structure + repo intelligence rather than raw model power.

Why it belongs here

A direct instance of agentic-coding-harness and agent-orchestration, and a sharp data point for the model-substrate thread: a good harness can make a cheaper model outperform a stronger one — structure as a substitute for capability. (Caveat: self-reported benchmark.)

Related

agentic-coding-harness · agent-orchestration · agent-skills · spec-driven-development · agent-kanban · claude-opus-4-8

Re-verified 2026-08-04 — the headline claim is now quantified

928★ / 108 forks, v6.0.1 (restores Cursor/Kiro install, drops a vulnerable js-yaml).

The claim this spoke has been chasing since June is still there and has grown numbers. The README now states: “Sonnet + agentsys produced more output with higher specificity than raw Opus — at 40% lower cost,” with a comparison table claiming 73–83% cost savings versus Opus alone at “the same outcome quality,” and the general conclusion that “with agentsys, model tier matters less.”

This is the spoke’s #1 growth edge restated by the vendor, not answered. It remains self-reported, with no stated task set, no trial count, no rubric and no independent replication — the exact shape jetbrains-ponytail-tested took an advertised −54%/−20% down to −15.4%/−10.3%. What changed since June is that the claim is now specific enough to be tested, which makes it a better target than it was: a controlled comparison varying model and harness would have a stated number to falsify.