A Hermes Agent skill that turns the main agent into a Coordinator dispatching a 4-role workforce: Researcher, Subject Analyst, Implementer, and Verifier. Ported from the eigent NPAS agent architecture. Platform-agnostic, model-agnostic, domain-agnostic.
You (main agent) → STRATA Coordinator
├── Researcher → SOTA scout, web research, multi-source verification
├── Subject Analyst → codebase mapper, 7-layer subject analysis
├── Implementer → ponytail ladder builder
└── Verifier → adversarial PASS/FIX/ESCALATE
A skill that slots into Hermes Agent. Load it (/skill strata), and the main agent becomes the Coordinator. It classifies every request into a lane (Single Agent / Research / Factory), shows a lane pill, and for Research lane dispatches a workforce via Hermes delegate_task. Each worker gets an XML-structured role prompt with explicit mandates, output schemas, and anti-patterns — copied verbatim from the skill's references/ directory.
Not an external engine. Not a separate process. No Docker, no API server, no model fine-tuning. It runs entirely inside Hermes using the tools that already exist: delegate_task, file read/write, and the MCP stack. The Coordinator IS the main Hermes agent reading this skill.
/skill strata
Then say anything research-shaped: "audit this codebase," "find the best scraping tool," "extract every user workflow from nuco.pro," "compare SOTA options for X."
The Coordinator classifies your request, shows the lane pill, dispatches workers, and synthesizes results into a dated folder at your working directory.
Every request gets classified into one of three lanes:
| Lane | When | What happens |
|---|---|---|
| Single Agent | Quick question, edit, <3 steps | You handle it. No workforce. |
| Research | "Find out", "compare", "audit", "extract workflows", "what does X actually do" | Full workforce pipeline (see below) |
| Factory | "Build / refactor / ship" a repo | Says "not implemented" and offers Research to spec it |
The lane shows as a pill: > Research | 2 subagents | [switch v]. User can override.
1. PLAN (Coordinator, <30s)
Decompose → decide worker count → create folder structure
2. RESEARCH + ANALYSIS (parallel fan-out)
Researcher → Secondary/agent_1_researcher_*.md
Subject Analyst → Secondary/agent_2_analyst_*.md
3. IMPLEMENT (sequential, after both return)
Implementer → Secondary/agent_3_implementer_*.md
Skip if research-only — Coordinator writes synthesis
4. VERIFY (sequential, after Implementer)
Verifier → Secondary/agent_4_verifier_*.md
On ESCALATE: re-dispatch with finding as constraint (max 2 rounds)
5. SYNTHESIZE (Coordinator)
Read all Secondary/*.md → Primary/final.md
Each worker gets a proprietary XML-structured prompt from references/. These are copied verbatim into delegate_task context — not summarized or paraphrased. The XML tags are load-bearing for GLM model reliability.
| Role | Does | Scale |
|---|---|---|
| Researcher | Multi-source web search (searxng), primary source reading (scrapling), SOTA scouting, cross-checking. Cites with timestamps. Labels every claim 【已确认】or 【推断】. | 1-N per research angle |
| Subject Analyst | 7-layer subject map of any target (codebase, config, schema, UI component tree): Inventory → Skeleton → Wiring → Load-bearing parts → Fragile parts → Negative space → Risk surface. "You map, you do not prescribe." | 1-N per codebase surface |
| Implementer | Builds/writes per Researcher brief + Analyst map. Ponytail ladder: stops at the first rung that works. Forbidden block: no backwards-compat shims, no premature abstractions, no tombstone comments. | 1 (or N for independent modules) |
| Verifier | Adversarial read-only. Re-reads every cited file:line. Checks for fabrication, missed SOTA, negative space gaps. Output: VERDICT (PASS | FIX | ESCALATE). "Praise is not your job. Errors are." | Always 1 |
Subagents inherit the parent session's model. Select GLM 5.2 from Hermes UI for heavy reasoning, DeepSeek V4 Pro Max for volume work. No per-agent model config — one less thing to break.
Every run lands in a dated folder:
<working directory>/
└── <Month YYYY>/
└── <Month Dayth>/
└── <Run Topic>/
├── Primary/ ← final deliverable(s)
├── Secondary/ ← subagent output files
└── Search-Scrape Runs/ ← raw dumps
Built from the eigent NPAS agent architecture (Eigent-source/backend/app/agent/prompt.py):
MULTI_MODAL_SYS_PROMPT→references/verifier.mdDOCUMENT_SYS_PROMPT→references/analyst.mdDEVELOPER_SYS_PROMPT→references/implementer.mdBROWSER_SYS_PROMPT→references/researcher.mdCOORDINATOR_SYS_PROMPT→references/coordinator.md
Every role prompt preserves eigent's XML tag structure, output schemas, collaboration patterns, and the CLAUDE.md writing habits (穴居语 compression, MUST/绝不 hard rules, 【已确认】/【推断】 labeling).
hermes skills install https://github.com/uhneer/strata-skill/blob/main/SKILL.mdOr manually:
cp -r strata-skill/ ~/.hermes/skills/software-development/strata/SKILL.md ← Coordinator playbook (loaded by /skill strata)
references/
├── coordinator.md ← Full Coordinator system prompt
├── researcher.md ← Researcher role (SOTA scout)
├── analyst.md ← Subject Analyst role (codebase mapper)
├── implementer.md ← Implementer role (ponytail builder)
├── verifier.md ← Verifier role (adversarial auditor)
├── dispatch_template.md ← Exact delegate_task call format
├── filepath_convention.md ← Dated folder structure rules
├── eigent-prompt-structure.md ← Reference: XML prompt architecture
└── codebase-ui-investigation.md ← Reference: fan-out patterns for codebase workflow extraction
- Domain-agnostic. No property-search guides, no visa tables, no relocation data. Those live in their own skills.
- No model IDs hardcoded. Model selection is the user's choice from Hermes UI.
- No infrastructure assumptions. Works wherever Hermes runs.
- P14 cleaner enforced on all web tool output before it enters any context.