Learn AI · Research

Research workflows, run from one command.

The most underrated part of an AI terminal: it is not a chatbot, it is a research lab. Here is what the built-in workflows do and when to use each.

Deep research, with citations

expert-ai deepresearch "the economics of modular construction"

The flagship workflow. It plans a multi-source investigation, gathers evidence, and produces a durable research brief with inline citations and a provenance sidecar — every claim marked as verified, unverified, inferred, or blocked. This is research-grade rigor, not a chat answer.

Literature reviews

expert-ai lit "test-time scaling laws"

Search papers and synthesize primary sources into a structured review. Good for survey chapters, grant background, and "what has anyone actually done here?" questions.

Paper audits & replication

CommandWhat it does
expert-ai audit <paper>Compares a paper's claims against its public code; flags mismatches and reproducibility risks.
expert-ai replicate <paper>Plans or executes a full replication workflow.
expert-ai review <artifact>Simulates a tough peer review with objections, severity, and a revision plan.
expert-ai compare <topic>Source-grounded comparison matrix with confidence levels.

Autonomous loops & watches

  • autoresearch — try an idea, measure results, keep what works, discard the rest, repeat.
  • watch — a recurring research watch on a topic, company, or paper area.
  • draft — turn findings into a polished paper-style draft with equations and citations.
  • code — full-spectrum software engineering and technical automation.

The 5-agent engine behind the workflows

Expert AI Terminal runs 5 specialized parallel agents that collaborate in one session — the core reason workflows can run autonomously:

AgentRole
ResearcherGathers evidence, reads sources, explores the web and your vault.
ReviewerChecks the work against criteria and catches gaps.
WriterDrafts the deliverable in a polished, structured form.
VerifierConfirms citations and provenance — marks claims verified/unverified/inferred.
ArchitectPlans the overall approach and keeps the session on track.

Add the 33+ specialized tools (file read/write, repo management, web search, sandboxed execution, Obsidian writes) and you have what ssv.asia calls "an end-to-end environment that turns a language model into a disciplined, reproducible engineering team".

Coding agents & harnesses (the 2026 wave)

2026's trend is agentic development: coding agents that can edit files and run commands, wrapped in secure harnesses that provide sandboxed runtimes, API gateways, and unified session state. The terminal is exactly this — ssv.asia calls it a "purpose-built agentic harness… it doesn't just chat — it reads and edits your local files, runs commands, and learns".

  • One-click launch — a .bat/.command launcher gets you a ready-to-run terminal.
  • Real-time transparency — see reasoning steps, approve or reject edits inline.
  • Integrated test bot — auto-creates unit tests, runs them sandboxed, reports pass/fail.
  • Policy-driven security — per-project rules like "no external SDK calls".
  • Persistent knowledge — style guides and library versions stay consistent across sessions.

Dive deeper on ssv.asia's 2026 AI trending agents & harnesses article ↗.

Why this matters for beginners

Most people start with chat and ask one question at a time. Workflows flip the model: you describe a goal, the agent runs a multi-step plan, and hands you a durable artifact you can cite, edit, and build on. You learn the craft of research, not just the craft of asking.

Provenance is the superpower. Every output that says "verified" carries a sidecar recording what was actually checked. That is what makes the difference between an AI answer and evidence you can trust.