Features¶
Everything HANI2 does, grouped by who uses it. Each entry says how to turn it on and where the details are.
For participants¶
The negotiation desk¶
| Feature | What it does | Details |
|---|---|---|
| Offer on the table | The partner's current offer, what it is worth to you, and Accept, Counter-offer and End buttons. Accepting and ending ask for confirmation. | The action dock |
| Counter-offer editor | One field per issue (buttons, lists or numbers), a gauge of the offer's value with a tick at your no-deal value, and an optional message. Large scenarios use three columns so the Send button stays in view. | During the negotiation |
| History thread | Every offer and message as a conversation, oldest first, scrolling to the newest. | The history thread |
| Your preferences in view | When a negotiation starts, your Preferences tool comes to the front (and the tools tray or tab opens on tablets and phones). | Your preferences when the negotiation starts |
| Clear names | With an assistant present, scenarios that call the sides human and agent show You and Partner instead. |
names |
| Questionnaires and consent | Optional questions after each negotiation and a consent form before the first. | Studies |
Layout¶
| Feature | What it does | Details |
|---|---|---|
| Three views | Full (tools left, offer generators beside the offer area), simple (one tools column) and focus (negotiation only). | Layout views |
| Resizable panes | Drag the dividers between panes; double-click to reset. Sizes survive a reload. | Resizing the panes |
| Enlarge a tool | Open any tool in a large window or full screen, and put it back where it came from. | Enlarge and put back |
| Tablets and phones | Landscape and portrait tablet layouts with a tools tray, and a phone layout with Negotiate / Tools / Assistant tabs. | Phones and tablets |
| Display settings | Light or dark theme, text size, history options, offers as sentences or chips. | Display settings |
Tools¶
| Feature | What it does | Details |
|---|---|---|
| Thirteen built-in tools | Preferences, Scenario Info, My Results, Value Histogram, Utility Plot, Negotiation Plot, Outcome Plot, Trace, Utility-based Selector, LLM, Random, All Results and Opponent Preferences. | Built-in tools |
| Negotiation Plot | The negotiation in utility space with the Pareto frontier, Nash and Kalai points and both offer trajectories (admins, or everyone if allowed). | Built-in tools |
| Any kind of preferences | The Preferences and Opponent Preferences tools show linear, affine, additive, generalized linear, polynomial, nonlinear, hyper-rectangle, weighted, discounted, rank-only, constant and probabilistic utility functions, and anything else as a ranked list. | Showing other preferences |
| Rearrange tools | Tabs, an enable switch, move between panes, hide and restore. | Tool panes |
| Smooth updates | Only visible tools are computed, in the background; charts draw when the browser is idle. | How tools are rendered |
The support agent¶
| Feature | What it does | Details |
|---|---|---|
| Assistant dock and status board | Chat with the assistant, answer its questions with buttons, and see its current advice on an always-visible board. | The assistant dock |
| Sixteen capabilities | Chat, notifications, tool control (enable, show, rearrange), fill the offer and message, send, accept, end, highlight, annotate, change the layout, custom panels, add tools and use tools. Admins grant them; each participant can switch them off. | Capabilities |
| Three autonomy levels | Suggest, semi-autonomous or fully autonomous for sending, accepting and ending. | Autonomy |
| Interface control | Spotlight buttons, issues and tools with a short note, attach notes, switch views, resize panes and show interactive panels. | Interface control |
| Finds and uses tools | Reads any tool as text (numbers, tables, chart summaries, inputs and buttons) and operates it like the participant, within its own permissions. A tool written for people needs no extra code. | Tools the agent can find, use and add |
| Adds tools while you negotiate | Puts its own tools, or registered ones, on the screen at the right moment; tools with timing = "agent" wait for it. |
Tools the agent can find, use and add |
| Your own agent | Subclass SupportAgent with or without a language model; ten tutorial levels from an echo bot to an agent that brings its own tools. |
Build a support agent |
| Autopilot | The support agent or any negotiator plays the human's seat. | Autopilot |
For competitions¶
| Feature | How | Details |
|---|---|---|
| Human-Agent Negotiation (no assistant) | hani2 --no-support-agent |
With or without a support agent |
| Hybrid Intelligence, built-in assistant | hani2 --support-agent on |
With or without a support agent |
| Hybrid Intelligence, a team's assistant | hani2 --support-agent team.py:TeamAgent (or package.module:Class) |
Submitting tools and agents |
| A team's tools | hani2 --tools team.py:TeamTool |
Submitting tools and agents |
| Opponent agents | hani2 --opponents BoulwareTBNegotiator,:template (also ?opponents= per link and opponents in env.json) |
Opponents |
| Saved competition setups | hani2 experiment create NAME --no-support-agent --opponents ... --tools ... --no-webmcp, then hani2 --experiment NAME |
Command line |
The opponent agent negotiates against the participant; the support agent helps the participant. Options, settings and admin screens use these two names throughout.
For researchers and administrators¶
| Feature | What it does | Details |
|---|---|---|
| Experiments | Self-contained folders with their own settings, scenarios, questionnaires and results. | Studies and results |
| Scenario sets | 19 scenario types grouped into sets (HAN2026, MIT2025, PieArena2026, ANL and one per ANL year, IAGO, HAN2027, HI2027, ANAC2027, or your own; sets can include other sets). Pick the set with hani2 --scenario-set, env.json or, as an admin, in the settings drawer or on the start screen. |
Scenario sets |
| Who plays which role | Rules for the participant's role: the scenario's suggested role, a fixed index, preferred role names, alternating or random. | Which side the participant plays |
| Your own scenarios | Write a story or a generator in Python and register it from a file, or serve saved scenarios from any folder with --scenarios. |
Writing your own scenario type |
| Recording and replay | Every negotiation is recorded completely — screens, clicks, pointer, keys, offers, assistant messages and function calls, language-model calls with tokens and cost — and can be replayed exactly as the participant saw it, with a check that the replay matches their screen. On by default. | Recording and replay |
| Evaluation metrics | Hard metrics (outcome, Pareto / Nash / Kalai distances, welfare, length, tokens and cost, move types, theory of mind, decision times, activity, use of the assistant, engagement flags), questionnaire scales (UEQ, TAM, opponent perception) and language-model judgements (assistant, opponent, human, assertions, error review), for any set of recordings; compare groups with scenario-balanced means and pass^k. hani2 replay evaluate, hani2 replay compare or the replay tool. |
Evaluation metrics |
| Evaluation questionnaire | UEQ, TAM, intention to use, feeling heard, and how the opponent came across, after every negotiation: HANI2_PER_NEG_YAML=evaluation. |
Subjective metrics |
| Results and events | A results CSV per participant (including the layout used: device and orientation) and an SQLite event log of every action, with analytics in the admin console. |
Results |
| Prolific | A no-login flow keyed by the Prolific ID, with practice and counted rounds and the completion link. | Prolific |
| Admin console | Live sessions (with each one's opponent and support agent), results, events, users and every setting, in the browser. | Admin console |
| Opening admin-only tools | Give any admin-only tool, or opponent utilities, to all participants. | Opening admin-only tools |
| Enlarged tools | Choose whether tools enlarge into a window, fill the screen, or not at all. | Enlarge and put back |
| Accounts | Password, GitHub and Google login, or none; hani2 users. |
Authentication |
| Language models | Ollama or any litellm provider, for the assistant and for LLM opponents. | Language models |
| WebMCP | Browser AI agents can read and (if allowed) act on the negotiation through the page; off for Prolific by default. | WebMCP |
For developers¶
| Feature | What it does | Details |
|---|---|---|
| Tools in Python | Write a class with a view() built from components; no frontend code. Register in env.json, with --tools, or as a package entry point. |
Build a tool |
| Tools agents can use | Give a tool an agent_hint, or override describe_for_agent(); timing = "agent" for tools only an agent adds. |
Make your tool usable by a support agent |
| Views for new kinds of preferences | Register a view for any utility-function class with a decorator or an entry point. | Showing other preferences |
| Headless testing | hani2.testing.make_session runs a whole session without a browser. |
Testing agents offline |
| Installation | pip install hani2 on Python 3.11 to 3.15, on Linux, macOS and Windows. |
Installation |