Personal AI Delegate (PAID)

An independently built "personal AI delegate" — user-authorized, user-accountable, information belongs to the user; the AI version of the principal-agent relationship in the same category as lawyer / accountant / doctor.

1. Core Product / Service

PAID is a personal delegate agent that runs in IM, letting the AI handle asynchronous collaboration with other people or other agents within scopes the user has authorized.

• Three-tier response modes: direct answer / request authorization / decline; initial human-machine ratio 50/50, target reaching 80%+ auto-answer.
• RAG input layer: locally authorized information (owner profile, persona, SOPs, counterparty notes) + web search tool combination.
• Multi-platform IM: Phase 1 doubles down on Slack, Phase 2 bridges across organizations, Phase 3 fans out to Lark / Telegram / Slack multi-platform.
• Local Dashboard: approval queue, outbound queue, audit log visualization (v0.5 is CLI for now, dashboard is a W2+ TBD).
• Stack: runtime runs on hermes-agent, with an OpenClaw-style IM gateway; model routing via openrouter.
• Deployed form: Hermes plugin paid-v1, MIT open-sourced from 2026-05-02.
• Four-layer pipeline: L1 input regex / L3 approval loop (slash command version, no card button) / L4a + L4b output regex (observer-only) / L4c LLM post-check and L4d source attribution are listed as W2 TBD.

2. Target Users & Pain Points

Primary persona: Seniors managing Juniors — founders / research leads / training mentors / senior professionals. The pain is being drowned in low-density queries ("how do I do this", "how did you handle that before"), yet unable to outsource directly because the cost of a wrong answer falls on them.

Relationship to ai-human-hybrid / human-in-the-loop-ai: the core mechanism is "auto within scope + escalate when out of bounds", not full autonomy and not pure copilot.

3. Competitive Landscape

Differentiation anchors: (1) runs in the user's existing IM rather than a new app; (2) explicit principal-agent legal metaphor rather than "AI assistant"; (3) output layer regex + LLM dual gate — out-of-bounds, prefer placeholder/decline over making things up.

4. Unique Observations

• Design tradeoffs: v0.5 cut the dashboard, card button, auto counterparty discovery, FTS5, and web retrieval, forcing the minimum closed loop on the J3 approval loop (slash command suffices) — prove the 3-state pipeline runs end-to-end before expanding.
• Key technical risk: whether relying on pre_llm_call to inject context so that the LLM proactively says placeholder / decline is reliable; if unreliable, fall back to a pre_gateway_dispatch hook that rewrites the message. At that point Layer 4a/4b must upgrade from observer to enforcer.
• Anti-consensus success criteria: success is measured by falsifiable usage indicators, not product perfection — "whether it has been pushed in front of users" over "whether it's done well".

5. Financials

• Funding: none; open-source project with no fundraising.
• Monetization path: free during the initial phase; the goal is adoption before any paid experiment.
• Infrastructure cost: models routed via openrouter; marginal cost is mainly LLM tokens.

6. People & Relationships

• Builder: solo independent developer
• Connected components: hermes-agent (runtime), review-agent (peer-review reference, same-stack sibling), openrouter (model routing)
• Related modules: claude-code-sessions (development process dependency), deep-research-workflow (info organization), ai-human-hybrid / human-in-the-loop-ai (design philosophy)
• Testers: small invite-only tester pool for first-batch feedback

Sources

• Internal research notes
• Internal research notes