If your agent improves when you add instructions, then suddenly gets worse, you may not have a prompting problem. You may have a harness problem.
We saw this firsthand building a finance research harness steering Kimi K2.6. We were about to announce during the week of the Opus 4.8 launch, riding a wide lead on the then-leading benchmark, Vals AI v1. Then Vals AI released v2 and Anthropic showcased Opus 4.8 on it — moving the spotlight to the harder benchmark overnight. On v2, our edge had collapsed to zero. Prompts had grown past 2,500 lines of instructions and 50+ tool definitions, many overlapping or conflicting. Adding more instructions stopped moving the needle. That's when Reef was born.
What you'll find in this repo: the framework itself (reef/), the finance harness that came out the other side at 82.6% on Vals AI Finance v2 and $0.13 per query (alphacumen/), the autonomous optimization-loop coordinator that drove +59pp on HotpotQA across 108 unattended LoRA experiments (polyp/), and the token-accounting diagnostic that started everything (claude-code-token-xray/) — all Apache 2.0.
Three independently-installable packages plus a diagnostic. Each has its own README, its own quickstart, and its own story.
| Path | What it is | Headline |
|---|---|---|
reef/ |
The harness framework. ReAct loop, skills-as-folders, declarative runtime constraints, direct-provider LLM client. Domain-agnostic, ~1,900 LOC. | The substrate everything else builds on. |
alphacumen/ |
Finance agent harness built on Reef. 7 specialists, 69 skills, the postprocessor synthesis path. | 82.6% on Vals AI Finance Agent v2 · 90% on Vals AI v1.1 · 89.3% on FinanceBench · $0.13/query |
polyp/ |
Autonomous optimization-loop coordinator. Postgres-backed state machine: Architect → Worker → Analyzer → Auto-suggester. Drives any try-evaluate-iterate problem; agent-driver-agnostic (Claude Code by default, your own Reef harness, or any coding agent). | +59pp on HotpotQA over 108 unattended LoRA fine-tuning experiments on gpt-oss-20b in 3 days. |
claude-code-token-xray/ |
The diagnostic that started the company. Breaks your ~/.claude logs into tokens, time, and cost. |
~29M unique tokens billed as 4.35B (~150×) — 84% of the bill is input. Nothing leaves your machine. |
Want to build your own domain agent? → reef/ — read the framework write-up, copy reef/examples/equities/, rewrite four pieces.
Want state-of-the-art finance answers right now? → alphacumen/ — pip install, set CORAL_API_KEY, ask. Runs against the hosted ~4.5 TB pre-processed finance corpus.
Want to run unattended optimization sweeps? → polyp/ — stand up Postgres, cbq init-db, point at a worker script. Read the LoRA trajectory writeup for a real-world run.
Curious where your Claude Code bill actually goes? → claude-code-token-xray/ — runs on your own logs in a few seconds. Full breakdown.
┌─────────────────────────────────────────┐
│ reef/ (framework) │
│ ReAct loop · skills · constraints │
│ provider-neutral LLM client │
└─────────────────────────────────────────┘
▲ ▲
│ imports │ drives (one option)
│ │
┌─────────┴──────────┐ ┌────────┴─────────┐
│ alphacumen/ │ │ polyp/ │
│ 7 specialists, │ │ Architect → │
│ 69 finance skills │ │ Worker → │
│ 82.6% Vals v2 │ │ Analyzer → │
│ │ │ Auto-suggester │
└────────────────────┘ └──────────────────┘
- AlphaCumen → Reef: hard import. Same
run_reactloop,@skill_fndispatch,llm.chatclient. - Polyp → Reef: optional. Polyp coordinates the loop; the agent driving each phase is yours — Claude Code by default, or wire a Reef harness if you want a domain-specialized Architect.
- Reef → nothing: zero finance, zero queue, zero opinion about your data plane. Vendor on its own.
reef/ # this repo
├── reef/ # the framework (agent harness primitives)
├── alphacumen/ # worked finance instance
├── polyp/ # autonomous optimization loop coordinator
└── claude-code-token-xray/ # the diagnostic that started it all
Each package owns its own pyproject.toml, README.md, and tests. Install only what you need.
Apache 2.0 — see LICENSE.
Hitesh Jain & Divy Vasal — Coral Bricks