Institution Mode

Build a configurable, read-only Roundtable of specialist agents for investment research and decision review

Institution Mode is Minara's workspace for composable multi-agent investment research. Instead of asking one model for a longer answer, you build a Roundtable: a sequence of analysis phases, the agents assigned to each phase, and the rules that govern how evidence becomes a final conclusion.

Institution Mode is read-only by design. It can research, debate, assess risk, and propose a plan. It cannot place an order or move funds. Execution always begins as a separate request and follows the normal preview and confirmation controls.

A Roundtable, not a fixed pipeline

A Roundtable has three layers:

Layer	What it controls
Roundtable	The complete workflow, its title, and its reusable template
Phase	One step in the workflow, including its purpose, prompt, type, rounds, and participants
Agent	One specialist's identity, research framework, instructions, skills, and operating budget

The default Roundtable offers a complete investment-committee process. You can use it unchanged or reshape it around the question. A company earnings review may need several independent accounting perspectives. A token review may add on-chain fundamentals and market-structure specialists. A macro allocation question may spend less time on technical analysis and more time on scenarios and cross-asset risk.

The default Roundtable

Stage	Purpose	Default participants
Retrospect	Bring forward relevant lessons from prior decisions	Historical Pattern Lead
Research	Collect independent evidence in parallel	Fundamentals, Sentiment, News and Macro, Technical
Research Debate	Test the strongest opposing cases	Bull and Bear researchers
Research Plan	Reconcile evidence into scenarios and priorities	Research Manager
Trading	Turn the view into a non-executing entry, sizing, stop, target, and horizon plan	Trader
Risk Debate	Stress-test the proposed exposure	Aggressive, Conservative, and Neutral risk roles
Investment Committee	Produce the final decision and report	Portfolio Manager

For a basket of assets, Minara runs the per-asset work and then adds an Allocator to compare candidates or propose cross-asset weights.

Compose the analysis phases

Every phase uses one of five execution types:

Type	Behavior	Typical use
`Context`	Establishes shared background before the main research	Prior decisions, mandate, portfolio constraints
`Parallel`	Runs several agents independently	Fundamentals, technicals, news, on-chain, industry research
`Debate`	Lets at least two agents challenge one another for one or more rounds	Bull versus Bear, thesis challenge, risk review
`Synthesis`	A manager combines contributor evidence into one view	Research plan, trade construction, scenario synthesis
`Final`	One manager produces the required final conclusion and report	Investment committee decision

Phases can be inserted, duplicated, deleted, and reordered. Each phase has its own title and common prompt, and Debate phases can set the number of rounds. The Final phase is fixed at the end so every valid Roundtable has one clear owner for the final report.

Customize each Agent

An Agent is more than a display name. The Roundtable Builder lets you configure:

Identity and role: name, role title, and research framework.
Instructions: start from a built-in template, customize its system prompt, or begin with a blank prompt. A compiled-prompt preview shows how Agent, phase, task, and output instructions fit together.
Skills: allow every eligible research skill or select a limited set. Only skills backed by read-only tools are eligible.
Web research: enable or disable web search for that Agent.
Work budget: choose a single pass or an iterative loop, then set iterations, tool-call limit, timeout, and output tokens per iteration.
Reusability: save the configured Agent as a template for another Roundtable.

This makes specialization explicit. Two Agents can inspect the same asset with different evidence standards, time horizons, or investment frameworks without blending their assumptions into one hidden prompt.

Templates at three levels

You can save and reuse:

an Agent template for one specialist;
a Phase template containing its prompt, type, and participants;
a Roundtable template containing the complete workflow.

Built-in templates remain unchanged. Loading a template creates a working copy for the current session. Saving the session makes that Roundtable the session's active configuration.

What a run records

During a run, the Roundtable view shows phase and Agent progress, tool activity, completed conclusions, and failures. A completed run stores an immutable snapshot of:

the Roundtable template and phase order;
every Agent's prompt mode, skill policy, web-search setting, and loop limits;
the skills and tool calls actually used;
per-phase status and each Agent's output;
the final result and generated HTML or PDF report files.

Later edits affect the next run only. They do not rewrite an earlier decision process. A run may finish as complete, partial, or insufficient_data; failed phases remain visible instead of being silently removed from the record.

Cost follows the design

A Roundtable with more Agents, debate rounds, iterations, and tool calls costs more and takes longer. The Builder estimates Agent count, iterations, tool calls, and maximum output tokens before you save. Treat that estimate as a design constraint: add a role when it contributes a distinct decision value, not simply to make the committee larger.

Safety boundaries

Roundtable Agents receive only eligible read-only research capabilities.
Trading, messaging, deployment, payment, and nested Institution capabilities are excluded from Agent skill scopes.
A Debate phase requires at least two Agents.
Synthesis and Final phases require exactly one manager.
A valid Roundtable requires exactly one Final phase, with one Agent, at the end.
The user's final decision and all fund-moving confirmations remain outside the Roundtable.

When to use Institution Mode

Use it when the quality of the decision benefits from distinct perspectives, explicit disagreement, and a reviewable path to the conclusion. For a price lookup or a simple fact, ordinary chat is faster. For a consequential single-asset decision, a basket comparison, or a custom research process that you expect to reuse, a Roundtable provides the stronger structure.