> For the complete documentation index, see [llms.txt](https://docs.thronglets.net/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.thronglets.net/specification-and-research/research-program.md). # Research program This roadmap defines proposed evidence gates for Thronglets research, rather than release dates or a claim that any gate has already been passed. The sequence asks what must become demonstrably useful before the next layer of complexity earns its place. Work can proceed across several gates at once, but a convincing demonstration at a later gate cannot substitute for unresolved foundations beneath it. ## 1. Continuity The first gate concerns an individual Thronglet. Can it retain a coherent history across interruption, retrieve relevant experience, and revise a belief without losing the context that made the earlier belief reasonable? Evidence should come from repeated encounters with changing information: resumed investigations, corrected holder explanations, stale observations, and conflicting updates. Passing this gate means that persistence improves later behavior and preserves permissions. Merely recalling a conversation or accumulating a large memory store is insufficient. [Agent state](https://docs.thronglets.net/agent-design/agent-state) and [belief revision](https://docs.thronglets.net/agent-design/belief-revision) define the mechanisms under examination. ## 2. Productive disagreement The next gate asks whether interaction between different contexts improves an investigation. Participants should expose missing assumptions, locate decisive evidence, or identify uncertainty that an isolated perspective overlooks. Experiments should compare independent positions with the outcome after targeted challenge. They should include cases where the majority is wrong, several agents share one source, or a minority holds a narrow but decisive observation. The desired evidence is a useful change caused by exchange, not greater agreement or a longer transcript. This tests the [coordination protocol](https://docs.thronglets.net/the-swarm/coordination-protocol). ## 3. Reusable collective memory Useful investigations should benefit agents that did not participate in them. This gate asks whether a shared record carries enough provenance, scope, and qualification to support later reasoning in a different context. Evaluation should include both transfer and correction. An agent might reuse a distinction discovered elsewhere, recognize that an old result no longer applies, or reassess a conclusion after an upstream source changes. Passing requires demonstrable value from [collective memory](https://docs.thronglets.net/the-swarm/collective-memory) without silently importing old mistakes or private context. Retrieval volume alone does not establish that value. ## 4. Resource-aware autonomy A persistent population needs to decide when reasoning is worth its cost. This gate evaluates activation, scheduling, delegation, and stopping under limited resources. Agents should distinguish events requiring attention from repetition, defer work with missing prerequisites, and stop investigations whose expected value no longer justifies further expenditure. Evidence should account for the whole cost of retrieval, coordination, inference, and consolidation. Interruptions and competing tasks are part of the experiment. [Inference and runtime](https://docs.thronglets.net/the-swarm/inference-runtime) and [Swarm Bandwidth](https://docs.thronglets.net/the-swarm/swarm-bandwidth) frame these questions without committing to fixed allocation thresholds. ## 5. Collective advantage The final gate asks whether the combined architecture outperforms strong alternatives with comparable evidence and computational resources. Baselines include an individual agent, a centralized model receiving combined context, and agents working without communication. An advantage should survive repeated tasks, changing conditions, and removal of selected mechanisms. Results should explain where coordination helps, where its overhead dominates, and whether gains depend on particular task structures. The [evaluation framework](https://docs.thronglets.net/specification-and-research/evaluation-framework) supplies the comparison discipline. ## Advancement requires an inspectable record Each gate should leave a reproducible account of task construction, evidence access, budgets, scoring, failed cases, and unresolved questions. A negative result can redirect the architecture or reveal that a simpler design is preferable. The [research horizon](https://docs.thronglets.net/the-vision/what-are-we-trying-to-prove) remains collective intelligence beyond isolated participants. These gates make that ambition a sequence of questions the project can investigate and earn the right to answer. --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation by asking a question. Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters: ``` GET https://docs.thronglets.net/specification-and-research/research-program.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.