agenthropic

Recommendation

⚠️ SUPERSEDED (2026-07-04, banner added 2026-07-06). The evidence and grades in this dossier remain valid and citable. The recommendation below (fork simple10; graft cast analytics.ts) is superseded: the project is a greenfield clean build per ../analysis/best-path-decision.md, and cast/disler/nirdiamant code is clean-room-only per CD-9. Current entry point: ../analysis/PROJECT-STATE-2026-07-06.md.

Status: recommendation only. No action taken. The direction is Ivan’s to pick.

Decision

Fork simple10/agents-observe as the base for agenthropic. This flips the vendor panel’s primary pick (hoangsonww → simple10).

Why simple10

• Highest independent grade (A−) and it also wins the panel’s own weighted model (4.1) once the refuted “no DAG” claim is corrected.
• Genuine subagent hierarchy graph — buildAgentTree() + live force-directed graph — the exact core capability agenthropic exists to provide.
• Real tests (78 files, 1,985 expect()), clean ports/adapters architecture, strategy-pattern agent classes → the most forkable base for an eventual OPCⁿ product.
• MIT + a real LICENSE file (© 2025 Joe Johnston) → no legal blocker, unlike disler (no license) or cast (license-by-badge).
• Docker is avoidable — AGENTS_OBSERVE_RUNTIME=local / just start-local runs pure Node + native better-sqlite3, so it lives under launchd next to Ollama and the Telegram bot with no always-on daemon (the trade-off the panel overweighted).

See projects/simple10.md for the full evidence.

Step plan (once approved — not started)

1. Harden first, before running anything networked.
   • Bind 127.0.0.1; add a DASHBOARD_TOKEN + same-origin WS check.
   • Run via AGENTS_OBSERVE_RUNTIME=local under launchd — no Docker daemon.
   • Redact stored tool payloads.
2. Validate the graph on a real session. Run one genuinely subagent-heavy orchestration and confirm buildAgentTree renders the nesting the way expected. This gate comes before any grafting — if the graph doesn’t hold up, reconsider.
3. Graft the two best rival ideas:
   • hoangsonww’s formatTelegram webhook provider (webhook-providers.js:177) → alerts to @baev_bot_bot.
   • cast’s delegation-savings metric (analytics.ts:233-310, ~50 LOC; re-verify the hardcoded pricing table).
4. Extend toward the gap (the OPCⁿ moat): persist the subagent edges as first-class rows (today they’re event-derived at render time); add a global, historical DAG view (ELK/Graphviz over the persisted tree); plan cross-machine aggregation.
5. Keep hoangsonww cloned as reference only — for its D3 Sankey / aggregate-DAG polish and its webhook schema.

Alternatives (documented, not recommended)

• If out-of-box richness today matters more than clean ownership: adopt hoangsonww loopback-only, delete run.js/run-spawner.js first (RCE), and accept the bus-factor-1 support burden on 92k LOC by a single author.
• Do not base anything commercial on disler (no license, no tests, no hierarchy) or cast (CAST-OS lock-in — 37/51 routes welded to getCastDb — + license-by-badge).
• Build-from-scratch stays on the table as a later option, but forking simple10 is strictly faster to a working DAG and inherits its test suite; a clean rebuild is only justified if the graft/extend work reveals simple10’s model is a poor fit.

What to steal, regardless of base

From	What	Where	Size
cast	controlGate.ts read-only-by-default security shape	controlGate.ts	~73 LOC
cast	delegation-savings (Haiku re-priced at Sonnet)	analytics.ts:233-310	~50 LOC
hoangsonww	Telegram/webhook provider + alert schema	webhook-providers.js:177	small
disler	the clean hook→HTTP→SQLite→WS ingest loop (as a teaching reference)	send_event.py	~180 LOC
nirdiamant	non-destructive git-stash+tag run-checkpoint pattern	snapshot handler	small
claude-code-templates	chokidar watch over ~/.claude JSONL; zero-install DX bar	AgentAnalyzer.js	pattern

Open questions for Ivan

1. Session-scoped vs global DAG — is a per-session tree enough at first, or is the global cross-session/cross-machine orchestration graph a v1 requirement?
2. Commercial intent (OPCⁿ) — does “may commercialize later” harden into a v1 constraint now (affects license hygiene, multi-tenant data model)?
3. Fork vs build — approve the simple10 fork, or want a scratch build evaluated first?

This site is open source. Improve this page.