MRS
Menu

Frontier research preview

Know what to learn, adopt, or test next.

Frontier turns a technical question into a sourced map of the concepts, claims, choices, and open tests that matter to the project you are trying to build.

Explore the live preview See the product direction

Product direction

From a research question to a project decision.

Pull is the first wedge: understand the relevant design space and decide what to learn, adopt, test, or ignore. Hook and Test close the loop later.

First product wedge

02 Position

Pull in frontier

See where your setup sits—and what is worth testing.

View the prototype
pull / evidence comparison3 named sources · fit pending
Illustrative project profile · no repository connected

A local coding-agent harness with a growing skill library.

Markdown skills load on demand. Routing is description-based. No behavioral admission gate is declared.

  1. 01
    UnitMarkdown skill document
  2. 02
    RegimeLocal-first · limited compute
  3. 03
    EvidenceDirect comparison ≠ constraint
Frozen corpus · named evidence

Related is not yet comparable.

apply project fit
  1. HarnessOpt-Bench Figure 5 showing crossing model lines across opencode, mini-swe-agent, and goose, plus native-harness differences by model.
    Figure 5 · p. 10 · cropped from paper
    External constraintRelated source

    HarnessOpt-Bench: Evaluating LLMs at Harness Optimization

    Varun Ursekar et al.

    arXiv:2608.06301v1 · CC BY 4.0The best harness depends on the model.
  2. Heatmap showing agent ranks changing across two task paths for average performance, forward transfer, backward transfer, and forgetting.
    Figure 1 · p. 1 · cropped from paper
    Direct comparisonRelated source

    PATH-Bench: Path-Dependent Evaluation of Lifelong Agents

    Xidong Yang et al.

    arXiv:2608.01149v1 · CC BY 4.0Agent ranks change across histories.
  3. Bar charts comparing default and two shuffled task orders for AWM and RBank on WebArena and VisualWebArena.
    Figure 3 · p. 6 · cropped from paper
    Direct comparisonRelated source

    On the Fragility of Self-Improving Agents: Variance, Task Order, and Underspecification

    Qinyuan Ye et al.

    arXiv:2608.18066v1 · CC BY 4.0Task order changes the measured result.
Project-relative disposition

No recommendation before filtering.

Three papers touch skill or harness improvement, but their units, regimes, and evidence roles differ.

Inspect setup, regimes, and deliverablessource boundary + exact versions

Setup used

Repository
Not linked · illustrative profile
Constraint
Local-first · limited compute
Current unit
Markdown skills loaded on demand
Unknown
Whether reuse survives repository and harness change

Proposed deliverables

  • Evaluation-control note: repeats + order perturbation
  • Local history-probe plan: proposed, not run
  • Skip receipt for the unmatched whole-harness campaign

No repository link, code change, run, or applied state is asserted.

Regimes + source receipts

On the Fragility of Self-Improving Agents: Variance, Task Order, and UnderspecificationDirect comparison · Two memory methods · three runs · default plus two shuffled orders

Qinyuan Ye, Yu Li, Yada Pruksachatkun, Jiaxin Zhang, and Chien-Sheng Wu

arXiv:2608.18066v1 · Figure 3 · p. 6 · CC BY 4.0 ↗
PATH-Bench: Path-Dependent Evaluation of Lifelong AgentsDirect comparison · Controlled histories · recurring probes · code and tool-use tasks

Xidong Yang, Xingyi Zhang, Wenhao Li, Wenyan Liu, Junjie Sheng, Yun Hua, Wei Yin, Tao Fang, Chuyun Shen, and Xiangfeng Wang

arXiv:2608.01149v1 · Figure 1 · p. 1 · CC BY 4.0 ↗
HarnessOpt-Bench: Evaluating LLMs at Harness OptimizationExternal constraint · Whole executable harness · protected holdout · four tasks · 111 scored runs

Varun Ursekar, Apaar Shanker, Yash Maurya, Shehab Yasser, Vijay S. Kalmath, Veronica Chatrath, and Yuan Xue

arXiv:2608.06301v1 · Figure 5 · p. 10 · CC BY 4.0 ↗
Corpus frozen 20 Aug 2026 · reviewed cross-paper evidence statuses; paper-local analyses remain candidate. Figure crops are unaltered apart from cropping and are attributed to their CC BY 4.0 sources.
Interactive product prototype. Double-click to inspect it.Open Pull deep dive

Next product loop

01 Watch

Hook into frontier

Know what moved—and what now contradicts your work.

View the prototype
01
Hookfrontier change
Frozen corpus20 Aug 2026 · 70 selected papers
04 / 04

The exact Figure 3 crop, paper identity, source span, and review boundary stay inspectable.

Step 4: Receipt opens. The exact Figure 3 crop, paper identity, source span, and review boundary stay inspectable.

Exact source receipt · candidate paper analysis

On the Fragility of Self-Improving Agents: Variance, Task Order, and Underspecification

Qinyuan Ye · Yu Li · Yada Pruksachatkun · Jiaxin Zhang · Chien-Sheng Wu

Open arXiv:2608.18066v1
Figure 3 from Ye and colleagues. WebArena pass rate is 54.8 for the baseline; default-order AWM and ReasoningBank are 54.1 and 56.3; Shuffle-1 falls to 49.1 and 49.8; Shuffle-2 reaches 49.8 and 50.7. A second panel reports VisualWebArena.
Figure 3 · p. 6 · rendered from the frozen local PDFYe et al., Figure 3. Cropped from arXiv:2608.18066v1 · CC BY 4.0
Project seam · NO REPOSITORY CONNECTEDThis update stops at the maintained question. Applying it requires a real project's task-order and run history.
Interactive product prototype. Double-click to inspect it.Open Hook deep dive

Later execution loop

03 Run

Test against frontier

Find out whether it works on your data and your harness.

View the prototype

Test · source-matched replay

Published order sensitivity becomes an 18-cell local plan.

Proposed · not runEvery local cell is planned
Source → matched design → blockers
Primary source · arXiv:2608.18066v1On the Fragility of Self-Improving AgentsYe et al. · 180-task WebArena GitLab subset
Published result · not an MRS runYe et al. · Figure 3
ReferenceWALT · no memory
AlternativeWALT + ReasoningBank
No memoryordinalseed42seed99
  1. Planned
  2. Planned
  3. Planned
  4. Planned
  5. Planned
  6. Planned
  7. Planned
  8. Planned
  9. Planned
ReasoningBankordinalseed42seed99
  1. Planned
  2. Planned
  3. Planned
  4. Planned
  5. Planned
  6. Planned
  7. Planned
  8. Planned
  9. Planned
18 cells · 3,240 episodesToken cap unsetUSD cap unsetmeaningful gain δ unset
Ye et al., Figure 3 · cropped from arXiv:2608.18066v1 · CC BY 4.0. The graph is published evidence; the matrix is a proposed MRS design and contains no local result.
Interactive product prototype. Double-click to inspect it.Open Test deep dive

The shared substrate

The graph is the record, not the interface.

Focused views answer a question.

A versioned evidence graph keeps every source, claim, concept, choice, and test inspectable underneath.

As sources or project assumptions change, the answer can be recomputed and diffed.

Release boundary

A useful preview, with the seams visible.

The public preview proves the research-space view. It does not pretend the full project-connected product already exists.

  1. NowResearch space

    A bounded corpus, reviewed synthesis, reading route, comparisons, and exact source excerpts.

  2. NextProject decision

    Add your goal and constraints; receive a sourced Learn / Adopt / Test / Ignore brief.

  3. LaterMaintained loop

    Monitor changes, recompute affected answers, and run source-matched tests in the local harness.

Human + machine

Machines compile the record. People decide what matters.

Agents search, extract, connect, and update. Researchers set the question, correct the model of the world, and decide what becomes part of the work.