Themis is an AI agent skill that converts a single general-purpose assistant into a structured, multi-agent research workflow. By establishing clear boundaries, it turns messy reasoning into an auditable process covering evidence retrieval, citation verification, adversarial review, and clean final reports.
Why Themis
While a standard LLM agent workflow can produce research-style outputs quickly, the underlying reasoning is often difficult to audit. Claims may lack backing, citation verification can be weak, and agents might agreeic-ally collapse too early.
Themis implements an agentic research workflow that makes structured research synthesis explicit. It forces the system to separate evidence retrieval from citation verification, maintain a claim ledger and a source register, and run peer review gates before compiling a clean final report.
How it works
Themis coordinates the research workroom using a structured orchestration manifest that drives the following stages:
Intake
↓
Workroom initialization
↓
Specialist role synthesis
↓
Independent review
↓
Claim ledger
↓
Evidence retrieval
↓
Retrieval gate
↓
Citation audit
↓
Adversarial challenge
↓
Peer review
↓
Public output gate
↓
Clean final report
Research roles
To manage this multi-agent research workflow, Themis defines four core management roles alongside 2–5 task-specific domain agents:
- Master Research: Oversees the workspace, manages the revision loop, approves dynamic specialist spawning, and compiles the final supervisor-grade final report.
- Evidence Retrieval Agent: Responsible for formulating search queries and gathering candidate sources for claims that require citations.
- Citation Auditor: Conducts metadata checks and citation audits to ensure the source supports the claims.
- Peer Reviewer: Audits methodology, conducts overclaiming reviews, and checks readiness for the public output gate.
Evidence and citation flow
Themis establishes a strict division between information gathering and validation:
- Retrieval: The Evidence Retrieval Agent queries host-provided search or database tools to compile candidates.
- Citation Verification: The Citation Auditor verifies candidate metadata and updates the claim ledger with verified or quarantined statuses.
- Peer Review: The Peer Reviewer conducts a source-grounded review to verify formatting, citation alignment, and check for overclaiming.
- Final Compilation: Once all gates pass, the Master Research agent writes the clean final report.
Repository layout
The installable skill package lives in:
skills/themis/
The .agents/ directory is maintainer-only automation and is not required by end users.
Install
Themis can be installed as a Codex skill, Claude Code skill, Gemini CLI skill, or Antigravity skill.
Using GitHub skill install:
gh skill install arsenelupin14/THEMIS themis --agent codex
gh skill install arsenelupin14/THEMIS themis --agent claude-code
gh skill install arsenelupin14/THEMIS themis --agent gemini
gh skill install arsenelupin14/THEMIS themis --agent antigravity
Manual install scripts are also provided:
./install/install-codex.sh
./install/install-claude-code.sh
./install/install-antigravity.sh
./install/install-opencode.sh
Usage
Ask your agent:
Use Themis to run a research workroom for this topic.
or:
Use Themis to audit this technical report for unsupported claims, citation issues, and overclaiming.
or:
Run a Themis workroom for this architecture plan and return a supervisor-grade final report.
Validation
Development setup:
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
Run validation:
pytest
python skills/themis/scripts/validate_themis_skill.py
python skills/themis/scripts/validate_research_mode.py
python skills/themis/scripts/validate_retrieval_contract.py
python skills/themis/scripts/validate_public_output.py tests/fixtures/valid-clean-final.md
python skills/themis/scripts/validate_claim_ledger.py tests/fixtures/valid-claim-ledger.md
python skills/themis/scripts/validate_orchestration_manifest.py tests/fixtures/valid-orchestration-manifest.yml
Safety
Inspect skills before installing.
Themis validation scripts are local, deterministic helpers that do not require network access. Retrieval tasks are performed by the host agent when search or database tools are provided.
License
MIT License.
No comments yet
Be the first to share your take.