The most common feedback on the [Draft] AI Governance Product Roadmap was that it tried to do too much. By exploring all the ways AI can enhance governance, we created an overwhelming list of ideas that left many unsure what to expect in the short term.
This final version narrows the focus to three core products, each corresponding to one of the three phases of AI governance. Other ideas from the first draft will manifest as future upgrades to these products, prioritized by user feedback.
Note: this roadmap defines the work of the product team housed within the Foundation’s Research team and funded by the Foundation. It is independent of any future roadmap for House of Stake, which will be defined by the incoming Head of Governance.
You can also find a Google Doc copy here, which you can download to markdown and discuss with your favorite AI assistant!
The road to superintelligent governance will progress in three overlapping phases:
-
Support: agents assist human decisions
-
Represent: agents proxy human decisions
-
Organize: agents coordinate human decisions
Each of these phases maps to a core product that we will focus on, in order:
-
Support → Proposal Dashboard (Q4 2025)
-
Represent → Delegate Agent (Q1 2026)
-
Organize → Manager Agent (Q2 2026)
Any other feature ideas will be focused through the lens of future upgrades to these core products, which we’ll prioritize based on user feedback.
Support: Proposal Dashboard (Q4 2025)
We’ll build a dashboard for AI-evaluated proposals, raising the bar on both sides by supporting both authors and reviewers. We’re focusing on a forum dashboard (off-chain) rather than a screening shade agent (on-chain), because the forum is the stage where low-quality proposals/responses create the highest attention cost.
-
Supporting reviewers: all proposals submitted to the forum will be screened using continuously improving evaluation criteria, helping reviewers to evaluate proposals. Note that this will be independent of the Screening Committee’s on-chain assessment.
-
Supporting authors: authors will also be able to screen their draft proposals pre-emptively against the same evaluation criteria then, if the proposal meets some quality threshold, publish their proposal to the forum with an “AI Screened” endorsement.
The evaluation criteria will be minimal to begin with:
-
Screening Criteria Pass/Fail: clarity, completeness, and adherence to proposal templates.
-
Alignment: alignment of proposal to Near’s mission, vision, values, and current strategic priorities.
The key upgrade paths that we anticipate are as follows, though these are subject to change and prioritization based on user feedback:
-
UX Enhancements: streamlining the Proposal Dashboard UX, e.g. copilot, conversational analysis, data filtering.
-
New Evaluation Criteria: adding new evaluation criteria – broad listening, community sentiment, author reputation (e.g. Near.social, Near Legion, NEARN), simulations, expert opinions (artificial and real), risk-benefit analysis, fact-checking claims, rough consensus, etc.
-
Evaluating Responses: expand the evaluation surface to include responses as well as proposals, e.g. rating response quality, highlighting noteworthy responses.
-
Evaluation Ecosystem: enabling users to submit new evaluation criteria, which other users can adopt into their own personal dashboards.
Represent: Delegate Agent (Q1 2026)
Instead of trying to build a toolkit for delegate agents, we’ll focus on building a single effective delegate which can be forked. Using the evaluation infrastructure we build for the Proposal Dashboard, we’ll create an AI Delegate shade agent that votes on the basis of the evaluations. As the Proposal Dashboard becomes more intelligent (see upgrade paths above), so too will the AI Delegate.
The key upgrade paths that we anticipate are as follows, though these are subject to change and prioritization based on user feedback:
-
Variations: fork the delegate into variations with more precise evaluation criteria that users may want to focus on (e.g. only vote on the basis of expert opinion).
-
Traceability: enable users to monitor the delegate’s voting activity, reasoning, and evaluate how effectively the delegate is representing them.
-
Feedback Loops: enable users to provide feedback to the agent (e.g. rating votes), so that it can refine how well it is representing its constituents.
-
Personalization: enable anyone to fork the delegate and personalize using private portable memory.
Organize: Manager Agent (Q2 2026)
If we build a shade agent capable of managing group membership (e.g. Sputnik DAO roles), this provides us a new primitive to scale the DAO. A Manager Agent will be able to add/remove members in real-time, on the basis of objective criteria, processing complex information, without ulterior motive; this will enable the DAO to delegate decision-making to localized groups without worrying about capture, scaling, or management failures (e.g. inadequate accountability). Note that this is not about agents making decisions, rather deciding who makes decisions.
The key upgrade paths that we anticipate are as follows, though these are subject to change and prioritization based on user feedback:
-
Specific Applications: tailoring the Manager Agent as needed to work for specific high-value groups, such as official House of Stake committees.
-
Management Logic: upgrading the logic available to the agent for appointing/removing members (e.g. elections, RFPs, reputation, accountability, conflicts of interest, talent/context gaps).
Acknowledgements
This roadmap was written by Jack Laing, as a synthesis of ideas and feedback from Lane Rettig, James Waugh, Klaus Brave, Cameron Dennis, Disruption Joe, Andrei Voinea, Eugene Leventhal, Puncar, Event Horizon, Nethermind, Hats Protocol, DeepGov, BlockScience, Metagov.