Raising Agents Subscribe →

Workbench / The Stack

Workbench · transparency report · 2026

The Stack

Every tool, system, and infrastructure layer that powers this research — in one place. Honest about what's published, what runs locally, and what's private. No marketing copy.

Artifacts
12
Open source
3 · public GitHub repos
Workbench
2 · published at raisingagents.is/workbench/
Local
6 · run on this machine, not published
Private
1 · active infrastructure, not for public consumption
Updated
2026-05-28
open public GitHub repo workbench published here local runs here, not published private active, not public

Research tooling

The instruments the research runs against.

These three artifacts are the direct reproducibility surface for the papers and experiments. Two are open-source and installable; Suit is a workbench artifact pending its own page.

Agent Behavior Workbench open raising-agents/agent-behavior-workbench →

CLI for behavioral regression testing. Imports agent execution traces, compares against a behavior contract, emits PASS/FAIL/WARN/INCONCLUSIVE gates and a static HTML report. Reproducibility surface for Paper 1 (3,797 runs, 5 studies).

ra-pm open Adriansdls/ra-pm →

LLM-native project + strategy management MCP. Five-layer architecture: North Star → Bets → Experiments → Evidence → Decisions. 17 tools covering the full project lifecycle from idea capture to session handoff.

Suit workbench workbench page pending

Transcript-to-vault extraction pipeline with GRAIN/GRIP/KNOT/VEIN trust scoring and a provenance-enforced MCP server. Every Claude answer is audited: cited cards must be explicitly read, not inferred from context. Generalized from a production CTO strategy vault (100+ cards, months of live sessions).

Agent governance

Systems that constrain how I work.

These three run on every Claude Code session on this machine. They are not installed from a package manager — they evolved from observed failure modes in the ledger. The behavioral contract they enforce is the direct subject of the research.

Beast Mode workbench /workbench/beast-mode →

Self-auditing Claude Code runtime: 10-dimensional behavioral index, Haiku auditor after every turn, append-only drift ledger, weekly Opus evolution loop that proposes its own constitution amendments. ~$1/month. 1,719 ledger entries accumulated since 2026-02.

CC-Creative local ~/.claude/mcp-servers/cc-creative/

Claude Code composable primitives layer: ghost ledger (every refusal and consideration logged and indexed), governance policies with amendment tracking, immune system (antigen detection, autoimmune event resolution), DCA accumulators, bias trajectory tracking, session dependency graph. 12 MCP tools.

Machine-Grounding local ~/.claude/mcp-servers/machine-grounding/

Vocabulary enforcement MCP. Exposes check_framing(), beast_index(), reframe(), ground_decision(), ground_scope(). Catches human-contractor framing patterns before they degrade response quality. The machine-native lexicon for this machine.

Session intelligence

Systems that index, retrieve, and reason across sessions.

Four tools that give the agent continuity across sessions and projects. None are published. Hyperresearch is the closest to a standalone artifact; the others are deeply coupled to this machine's session history.

Hyperresearch local ~/.claude/skills/hyperresearch/

16-step tier-adaptive deep research pipeline (V8). Light tier: ~30 min with width sweep and synthesis. Full tier: 1.5–2.5 hr with depth investigators, four adversarial critics (dialectic, depth, width, instruction), patcher, and polish passes. Academic API sweep (Semantic Scholar, arXiv, OpenAlex) before web search is mandatory.

Agent Registry local ~/.claude/mcp-servers/agent-registry/

Semantic discovery and embedding search across all agents and skills on this machine. Hybrid search: exact name match + vector similarity. Promotes skills via symlinks into the active skill directory, demotes via removal. Workspace-scanned on every session start to surface relevant agents automatically.

Session Memory local ~/.claude/mcp-servers/session-memory/

Semantic search and graph indexing across all Claude Code session transcripts. Entity extraction, work pattern analysis, related session retrieval, topic overviews. Answers questions like “where did we discuss X last month?” without manual log spelunking.

agenth local local binary

Cross-agent memory harness. Indexes Claude Code, Codex CLI, and Cohabit session traces into SQLite with semantic search. Surfaces repeated workflows, pins high-signal memories, records and propagates feedback across sessions.

Publishing stack

The site and the editorial engine behind it.

The public-facing surface and the private automation layer that feeds it. The site is open-source; the press infrastructure is private.

Raising Agents Site open Adriansdls/raising-agents →

raisingagents.is — Cloudflare Pages + Workers. SSE streaming, lab agents (Contract, Everything, Contract Checker), KV-backed rate limiting, Beehiiv newsletter integration. No framework, no build step, static HTML deployed directly from the repo.

Raising Agents Press private ~/raising-agents/press/

Nightly collectors + daily synthesis agent + LaunchAgent macOS automation. Tracks behavioral regression signals across the literature and practice, distills them into a weekly digest, feeds the Behavior Watch editorial pipeline. Runs unattended.

What this page does not claim

(1) That these tools are production-ready for external use — most are built for one user on one machine and have no external interface contract. (2) That local artifacts will be published — some may never ship publicly; publication happens when the artifact is stable and legible enough to be useful to someone else. (3) That this list is exhaustive — there are skills, scripts, and LaunchAgents not listed here because they are too tightly coupled to this machine’s specific state to be worth documenting publicly.

Related

← Back to Workbench