A visual, hands-on guide · June 2026

AI agents that don't just answer — they do the work.

How free and open tools like Hermes Agent, Google Antigravity and OpenCode can automate real legal and office tasks — working across WhatsApp, Telegram and email, around the clock — and why a new kind of “quant lawyer” matters.

5.6 cr
cases pending in India — the backlog AI can help clear
40%
faster on professional writing (controlled study)
20+
channels one Hermes agent can work across
24/7
always-on: scheduled & triggered tasks

The big idea in 60 seconds

From “chatbot” to “agent”

A chatbot replies. An agent takes a goal, makes a plan, uses tools (search the web, read your files, send a message, run code), checks its work, and reports back — and it can keep doing this on a schedule or when something happens.

🎯 You give a goal
in plain words
🧠 Agent plans
breaks it into steps
🛠️ Uses tools
files, web, messages, code
🧑‍⚖️ You verify
human signs off
✅ Done
output / notification
Keep this one rule in mind throughout 👉
AI drafts. A human decides. Every example below keeps a person in the loop to check the result before it counts.
Meet the tools

Pick the right tool for the task

All of these are free or open-source. You “bring your own” AI model — even a private one running on your own computer.

🤖

Hermes Agent

An always-on assistant that lives on your chat apps and runs scheduled jobs. Learns by writing its own “skills.”

Open-source (MIT)Runs locally

Best for: 24/7 monitoring, reminders, WhatsApp/Telegram/email helpers.

🛰️

Google Antigravity

An “agent workbench” that builds and tests things for you across an editor, terminal and browser — and shows its work.

Free previewCloud

Best for: building small tools/apps, multi-step projects with proof.

⌨️

OpenCode

A terminal assistant that can read a folder of files and do the task — and can run fully offline for privacy.

Open-source (MIT)Runs locally

Best for: document/data jobs, private/confidential work.

🔧

Cline · Aider · Goose

Friendly open helpers: Cline (in your editor), Aider (saves every change to git), Goose (reusable recipes).

Open-sourceLocal-friendly

Best for: repeatable team workflows you can share.

💬

Claude / Gemini / ChatGPT

The “brains” (models) the agents above can use. Some, like Claude, ship ready-made legal “Skills.”

Free + paid tiersCloud

Best for: the reasoning engine inside your agent.

🧩

Skills + MCP

Open “plug-ins”: Skills teach an agent a procedure; MCP connects it to your tools & data. Both open standards.

Open standards

Best for: owning your know-how instead of renting an app.

Why open tools? You keep control of your data, you're not locked into one vendor, and your team builds real skill. The honest trade-off: more setup and upkeep than a paid app. (More in Safety and Numbers.)

Spotlight

Hermes Agent — your 24/7 assistant on every chat app

Hermes is one agent with one memory that you can talk to on Telegram, WhatsApp, email and more. It runs in the background all day, does jobs on a schedule, and gets better by writing its own reusable skills.

📲Telegram 🟢WhatsApp 📧Email 💬Slack ✉️SMS 🔒Signal 🎮Discord 🤖 Hermes

One agent · one memory · 20+ channels CONFIRMED

Three powers worth knowing

Always-on & scheduled. Install it once as a background service; tell it jobs in plain English. CONFIRMED

# talk to it like a person: “Every morning at 9am, check the cause-list email, summarise new listings, and message me on Telegram.”

🧠 It writes its own skills. After you walk it through a task, it saves a reusable “skill” file and improves it over time — so next time it just does it. CONFIRMED (this is Hermes' real stand-out feature)

🔒 Private by default. Runs on your machine and can use a local model, so sensitive text need not leave the building. REPORTED

Honest note
The messaging + 24/7 + scheduling powers are a neat, one-command packaging of standard technology (chat-app bot APIs, webhooks, a scheduler). The genuinely distinctive part is the self-writing-skills + memory learning loop. Popularity claims (“most-used agent”, big star counts) are the maker's own marketing — treat as unverified.
# install (mac/Linux/WSL) curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash # then connect a channel and go always-on hermes gateway setup # add your Telegram bot token hermes gateway install # run 24/7 in the background hermes cron create "0 9 * * *" "Email me a digest of new orders"
How automation works

One mental model for every example

Every automation below is the same shape. Learn it once.

① Trigger
a message arrives, a schedule fires, or an event happens
② Agent
model + memory + tools + MCP make a plan and act
③ Action
read a file, search, draft, run code, fetch data
④ Human checks
a person reviews & approves
⑤ Output
the document, or a notification on your phone
🔔 Trigger by message

“What's my case status?” on WhatsApp.

⏰ Trigger by schedule

Every day at 9am, build a briefing.

⚡ Trigger by event

A new filing lands in the inbox.

The heart of it

Real-world examples — tasks & sub-tasks, automated

Click any card to open the step-by-step flow. Each shows the trigger, what the agent does, the tools/channels, and who checks the result.

All ⚖️ Legal & court 🏢 General office
📲Litigant asks case status on WhatsApp — gets a plain-language answer REAL · India

Setting: Adalat AI's WhatsApp helpline (launched 2026, built on Claude).

📲 WhatsApp messagepick language identify casefetch public e-Courts data summarise + translate orderreply + set alerts

Sub-tasks automated: language detection · case lookup · plain-language summary of the latest order · English→regional-language translation · follow-up Q&A · hearing-date alerts.

Tool / channel: agent on WhatsApp · public court data  |  Who checks: official court record is the source of truth.

Watch the cause-list 24/7 → ping the clerk when something changes REAL pattern · e-Courts

Setting: the e-Courts “Email My Case Status” + Supreme Court / Kerala HC WhatsApp alerts — the classic monitor-and-notify flow. With Hermes you can build your own.

⏰ runs all daycheck inbox/portal for new listings summarise what changed📲 notify staff on Telegram/WhatsApp

Sub-tasks automated: polling for updates · detecting new orders/defects/listings · summarising · routing the alert to the right person.

Tool / channel: Hermes cron + email/WhatsApp  |  Who checks: clerk reads the summary before acting.

📄Summarise a 40-page contract into 5 clear points with sources REAL capability
📄 drop in PDFOCR if scanned extract parties, dates, obligations, risksplain-language summary + page citations🧑‍⚖️ lawyer verifies

Sub-tasks: text extraction · key-term spotting · risk flags · each point linked to a page.

Tool: NotebookLM, or OpenCode/Goose + a PDF tool (private, local).

📊Turn a folder of contracts into a clause comparison table REAL capability
📁 folder of agreementsdefine fields (termination, indemnity, renewal…) agent reads eachone row per contract + page citations📊 spreadsheet

Sub-tasks: field definition · per-document extraction · citation linking · reusable saved agent for the next deal.

Tool: Cline/Goose + filesystem & PDF via MCP — runs over a local folder.

🗂️Build a case chronology & claim chart from discovery REAL capability
🗂️ files, emails, recordsextract dates, actors, facts normalise + cite pagesflag conflictsmap facts to issues

Output: a chronology, a timeline, and an issue map — exportable to Word/PDF.

💬A court self-help bot that answers the public (and checks itself) REAL · Nevada, USA
💬 public asks a questionsearch court-approved knowledge base (RAG) answer in 50+ languagesguided interview🧑‍⚖️ staff review transcripts

Why it's safe: answers come only from approved sources, and legal-aid staff review chats to patch gaps.

📥Inbox triage & smart auto-replies (private, on your machine) TEMPLATE
📥 new email arrivessummariseclassify & route draft reply from your knowledge base (RAG)📲 Telegram alert to approve

Sub-tasks: summarise · categorise · retrieve the right info · draft · notify. A local version uses Ollama so nothing leaves your computer.

Tool: Hermes, or an n8n template + Ollama + Telegram.

🌅A daily “morning briefing” delivered to your phone at 8am Hermes
⏰ 8:00am every dayscan overnight email + sources rank by priority + extract action items📲 send digest on Telegram

Set it in one sentence: “Every morning at 8, summarise overnight email and send me the top 5 things to do, on Telegram.”

📡Watch for new rules/notifications and flag what affects you TEMPLATE
⏰ daily schedulescrape gov / regulator sites assess impact & urgencyassign owner + deadline📧 prioritised alert

Great for: compliance teams and registries tracking UGC/DU/government circulars.

🔎Ask questions across all your own documents — privately REAL · open-source
📚 load your PDFs/Docssplit & index (on your machine) ask a questionanswer with citations to the source

Why it matters: a searchable, citeable “memory” of your office — built with open tools (LangChain + a local model), so confidential files stay in-house.

REAL = documented deployment/capability · TEMPLATE = published, runnable template (not a named end-user) · Hermes = native Hermes feature. Vendor time-savings are claims, not audited facts.

What to automate

Task → sub-task automation map

Filter by area. Colour shows how far it can go today: Full‑ish light check · Partial agent drafts, you finish · Human decides agent assists only.

All⚖️ Legal🏢 Office
CategoryExample sub-tasksHow farBest-fit toolWatch out for
Legal researchfind authorities, synthesise, check citationsHuman decidesNotebookLM / RAGfabricated cases
Draftingletters, notices, first-draft briefsPartialClaude / Geminitone, accuracy
Document reviewcontracts, NDAs, due diligencePartialCline/Goose + MCPmissed clauses
Case monitoringwatch cause-lists, deadlines, alertsFull‑ishHermes (24/7)wrong data
Litigation supportchronologies, claim charts, deposition prepPartialRAG + agentssource fidelity
Court admincause lists, transcription, translationPartialAdalat-style / SUVASconfidentiality
Intake & Q&Afront-desk questions, self-helpHuman decidesRAG chatbotwrong advice
Inbox & commstriage, summarise, draft repliesPartialHermes / n8nover-auto-reply
Daily briefingsscan, rank, send a digestFull‑ishHermes cronnoise
Monitoringregulations, websites, foldersFull‑ishscheduled agentstale sources
Knowledge baseask across your own files (RAG)Partiallocal RAGaccess control
Data wranglingclean, dedupe, convert filesFull‑ishOpenCodesilent errors
The numbers — honestly

Real gains, with a real warning

AI saves serious time on the right tasks. But on complex work an expert already knows well, it can even slow you down — so we measure honestly and always verify.

⏱️ Time impact by task (controlled studies)

+40% writing Writing +34% novices Support +14% avg Support avg −19% experts on code they know well (METR)

Sources: Noy & Zhang 2023; Brynjolfsson et al.; METR 2025. Faster ≠ always — context matters.

💸 Cost to run, per lawyer / month

Premium legal SaaS $200–$1,200+ Open / API per unit of work ~10× cheaper* *before your own setup & upkeep (the honest catch)

Per-seat SaaS prices are third-party estimates; open/self-host shifts cost to people & maintenance.

~44%

of legal tasks are exposed to automation (Goldman 2023) — a ceiling, not a guarantee.

17–33%

hallucination rate even in legal-grade AI research tools (Stanford) — so verify, always.

>40%

of agent projects may be cancelled by 2027 (Gartner) — skills & governance decide success.

Do it safely

Five habits that keep you out of trouble

Do ✔

  • Verify every output before it counts — click the citations.
  • Keep confidential data local / self-hosted (or on the institution's account).
  • Use public/official documents for demos and shared bots.
  • Start with one low-risk task; measure honestly.
  • Write down what AI may do, must be checked, and must never be automated.

Don't ✘

  • Paste student/client personal data into free consumer chatbots.
  • File or send anything you haven't read and checked.
  • Trust names, dates, fees or case-law without confirming.
  • Let AI make the decision — it drafts; a human decides.
  • Assume “legal-grade” means error-free (it isn't).
⚠️ The verification duty cuts both ways
Courts have sanctioned lawyers for AI-invented citations (e.g. Mata v. Avianca; recent Indian Delhi HC, Supreme Court and ITAT incidents). India's draft AI-in-courts rules and the EU AI Act both say: AI may assist, never adjudicate. A human stays accountable.
Why it matters

The “quant lawyer” & the case for an AI-Law course

A quant lawyer — like the “quants” who reshaped finance — is a legally trained person who can also command AI: write a Skill, connect a tool, build a private RAG, supervise the result with a lawyer's judgement. Not everyone becomes an engineer; the institution builds a spectrum of fluency.

Level 1
Prompt-literate
uses AI well & verifies it — the new baseline
Level 2
Skill-builder
writes reusable Skills & checklists
Level 3
Orchestrator
connects tools via MCP, runs multi-step agents
Level 4
Tool-builder
builds & self-hosts; owns the “quant” analytics
Why a course — and why now
Understanding AI's limits is becoming a competence duty (ABA Model Rule 1.1 cmt 8; Opinion 512). ~55% of law schools already offer AI courses. The cost of not teaching it shows up in every fake-citation sanction.
A deskilling-aware curriculum
Phase 1 — reason first: doctrine & drafting without AI, so judgement is built. Phase 2 — AI as assistant: agents with mandatory verification. Phase 3 — build: students author a Skill, wire MCP, ship one small supervised tool.
Start this week

Three small steps

1️⃣

Pick one task

Choose a weekly chore — say, summarising circulars — and do it once with NotebookLM or OpenCode.

2️⃣

Stand up one helper

Install Hermes, connect Telegram, and schedule a single daily briefing.

3️⃣

Draft one letter

Use an AI to draft a routine notice — then edit, verify, and send.

⏳ Small habit, big payoff
Saving 20 minutes a day is about two working weeks back over a year — time you redirect to judgement and people.