Most agent code is a prompt in a while-loop. I'm interested in the part underneath: what happens when the run dies halfway, when the context outlives the session, when a tool call has to be able to be wrong safely.

flowchart TB
subgraph runtime [" Runtime "]
BG["blackgeorge<br/><i>tool-calling, orchestration</i>"]
end
subgraph agents [" Agents "]
DSR["ds-review<br/><i>PR review</i>"]
SH["shandu<br/><i>deep research</i>"]
SRC["sourcery<br/><i>extraction</i>"]
end
subgraph memory [" Memory "]
AD["anchor-db<br/><i>code-local context over MCP</i>"]
end
subgraph infra [" Infrastructure "]
CW["crawlwall"]
HW["holloway"]
CSH["cashet"]
end
BG --> DSR

blackgeorgePython
Agent framework for LLM tool-calling and multi-agent orchestration. The layer ds-review runs on.

ds-reviewPython
Multi-agent PR reviewer. Starts from the diff, pulls the related code context, and comments only on findings tied to changed lines. Runs on DeepSeek for the 1M-token window, which is what makes whole-repo context practical rather than theoretical.

shanduPython
Deep research agent. Multi-agent search and scraping with source-credibility scoring, producing citation-backed reports. Model-agnostic through LiteLLM.

sourceryPython
Structured extraction from documents. Define a Pydantic schema, get typed entities back with exact source spans. PDFs, HTML, long text, OCR.

anchor-dbGo
Durable code-local memory for agents. Context lives in the Git repo beside the code it describes, served over MCP, CLI, HTTP, and a browser UI, so the reasoning behind a decision survives the session that produced it.

crawlwallGo
Caddy module for AI crawler access control. Bot verification, rate limiting, shadow-mode rollout, SQLite ledgers, signed crawl receipts.

hollowayGo
Self-hosted webhook relay. Every webhook persists to SQLite before delivery, so nothing is lost when the client is offline. Live dashboard, one-click replay.

cashetPython
Content-addressable memoization cache. Redis-backed, async, inspectable from the CLI.