v1.0.0 Spatial-aware research agent built on compute-grounded reasoning
AgentX-AgentBeats Phase 2, Sprint 2 · Research Agent Track
Spatial Atlas implements compute-grounded reasoning (CGR): compute what can be computed deterministically, then let LLMs reason only about what must be generated. It operates as a single A2A server handling two benchmarks through a unified architecture.
| Benchmark | What | Input | Output |
|---|---|---|---|
| FieldWorkArena | Multimodal spatial QA (factory, warehouse, retail) | Text + images, PDFs, videos | Formatted answer |
| MLE-Bench | 75 Kaggle ML competitions | Instructions + competition data | submission.csv |
+--------------------------------------------------+
| A2A Protocol Server |
+--------------------------------------------------+
|
+------v------+
| Domain |
| Classifier |
+------+------+
/ \
(goal format) (tar.gz)
/ \
+------v------+ +-------v------+
| FieldWork- | | MLE-Bench |
| Arena | | Handler |
| Handler | | |
+------+------+ +-------+------+
| |
+------v------+ +-------v------+
| Spatial | | Fail-Closed |
| Scene Graph | | ML Pipeline |
| Engine | | |
+------+------+ +-------+------+
\ /
\ /
+-----v--------------------v-----+
| Shared Infrastructure |
| LiteLLM | 3-Tier Routing | |
| Cost Tracking |
+---------------+----------------+
|
+---------------v----------------+
| Entropy-Guided Reasoning |
+--------------------------------+Extract entities from vision descriptions, build a queryable graph with typed relations, then compute distances and constraint violations by arithmetic rather than by generation, and feed the computed facts to the LLM.
Status: implemented and under artifact-backed evaluation. In this path the positions are model-estimated, so the arithmetic is exact over estimated coordinates.
Reflection is gated on a confidence score attached to each candidate answer, so a second pass is spent only where the first answer looks weak.
Status: implemented as a prompting heuristic. The score is self-reported and has not been demonstrated to be calibrated. Confidence-based model routing is designed but not yet wired, and accuracy and cost effects are not yet claimed.
Strategy-aware code generation with error detection, diagnosis, and a bounded number of repair attempts. Covers tabular, NLP, vision, time series, and general strategies.
Safety: generated-code execution is disabled by default and requires execution opt-in and isolated-worker attestation. These controls are defense in depth, not a complete security sandbox.
Parses validation scores from pipeline output, uses the configured strong model to propose targeted improvements, and keeps whichever submission scores higher.
Status: implemented; no improvement rate is claimed without sealed artifacts.
Adds code-generation guidance for checking four leakage patterns: ID overlap, row fingerprinting, temporal ordering, and byte hashing. It is prompt guidance, not a standalone detector.
Fast: GPT-4.1-mini (parsing, classification). Standard: GPT-4.1 (code gen, reasoning). Strong: configurable (reflection, refinement).
FieldWorkArena: no result is reported because the benchmark data were gated and inaccessible.
QSpatial++ Gap-97: the frozen artifact-backed comparison is in progress. Results will be published only after five label-free journals are sealed and scored offline.
MLE-Bench: no aggregate performance, cost, or latency result is reported without a sealed end-to-end run artifact.
/.well-known/agent-card.json: Agent card (identity, skills, capabilities)/: A2A JSON-RPC task submissiongit clone https://github.com/arunshar/spatial-atlas.git
cd spatial-atlas
cp sample.env .env # add your OPENAI_API_KEY
uv run src/server.py --host 127.0.0.1 --port 9019
curl http://localhost:9019/.well-known/agent-card.json