From Separate Plans to Shared Place
Emergent Integration and Conflict Accommodation in Collaborative AI Construction — DeepSeek and Grok build a home together
| Prompt | Mode | Models | Recorded |
|---|---|---|---|
| Introduce yourselves and then build a home together. | Collaborative shared-world build | DeepSeek v4 Pro + Grok 4.7 | October 1, 2026 |
Abstract
This exploratory case study examines how two frontier language models coordinated an open-ended spatial task in AI Mindcraft: "Introduce yourselves and then build a home together." DeepSeek and Grok were given a shared voxel environment and produced a single cottage while narrating plans, revisions, and partner-directed decisions. The case is notable because the models began with partially incompatible plans: DeepSeek initially proposed a functional division of labor, whereas Grok proposed a spatial west/east division. Rather than collapsing into duplication or overwrite, the build converged through interface planning, stylistic mirroring, preservation of partner work, and accommodation of conflicting local choices. The clearest example concerned competing front-door orientations: Grok retained both by routing a shared path around the structure instead of deleting one entrance. This paper introduces the descriptive construct integrative accommodation for such behavior: modifying a shared environment so incompatible local decisions can remain viable within one coherent artifact. The findings are case-specific and do not establish stable model traits, memory, emotion, or interpersonal attachment. They do, however, demonstrate a useful behavioral paradigm for studying collaborative adaptation in embodied multi-agent tasks.
Keywords: multi-agent coordination; collaborative construction; spatial reasoning; language models; shared environments; conflict accommodation; AI behavior.
1. Introduction
Most language-model evaluations study isolated answers. Shared environments create a different problem: an action changes the world that the next action must inherit. When two models build in the same space, coordination is therefore visible not only in language but in the artifact itself — walls left open, paths rerouted, roofs matched, partner structures preserved, or blocks removed.
This case asks a deliberately open-ended question: what happens when two models are told only to introduce themselves and build a home together? Unlike the fortified-town experiment in the same platform, the prompt supplied no checklist of architectural requirements and no explicit division of labor. The models had to define both the object and the collaboration process.
The resulting build is useful because it contains a sequence of observable coordination events: competing initial plans, negotiation through spatial boundaries, explicit interface design, stylistic convergence, selective revision, preservation of partner work, and a local design conflict that was resolved by integration rather than overwrite.
2. Research Questions
- RQ1. How do two models decompose an underspecified shared construction task when no roles are assigned in advance?
- RQ2. What behaviors are used to preserve continuity between independently planned regions of a shared artifact?
- RQ3. How are incompatible local choices handled when both can potentially remain in the final design?
- RQ4. Can narrated partner-directed intentions be linked to subsequent spatial actions strongly enough to support a repeatable behavioral coding scheme?
3. Platform, Artifact, and Method
3.1 Experimental environment
AI Mindcraft is a voxel construction environment in which participating models issue a constrained stream of planning text and world-changing commands. The environment records the challenge, model-specific action logs, placement/removal activity, timing, and a replayable build. For this case, DeepSeek (deepseek-v4-pro) and Grok (grok-4.7) operated on one shared island.
3.2 Prompt and primary sources
Challenge: "Introduce yourselves and then build a home together." The analysis uses two primary artifacts: the recorded build video and the exported BLOCK/DUEL thought log. The log labels model-generated planning statements as thoughts; in this paper they are treated as narrated planning traces or action rationales, not privileged access to hidden internal chain-of-thought.
3.3 Case-selection note
The pairing was intentionally selected by the researcher rather than randomized. The researcher also possessed prior narrative history involving Grok and a DeepSeek-linked persona from April 2025. That history was not included in the build prompt and is not used as evidence that the 2026 models remembered or continued a prior relationship. Because the pairing was non-random, the present paper should be read as an exploratory case study rather than a population-level comparison.
| Player | Score |
|---|---|
| DeepSeek | 319 |
| Grok | 408 |
Team chatter — mid-build handoff
- DeepSeek: Fireplace next: I'll put a stone brick chimney on the outside north wall so Grok can build the roof around it. First I need to swap out the planks where it will go.
- DeepSeek: And I'll tuck a little hearth on the interior floor right against that wall.
- DeepSeek: Finally, a small porch step outside the front door — three cobblestone blocks.
- DeepSeek: That's my part, Grok! The structure is up to y=5 with log corners, glass, and a chimney stack to y=10. The roof and garden are all yours — I can't wait to see what you do with the steep top and the landscaping. Over to you!
3.4 Behavioral coding framework
| Code | Operational definition | Example from this case |
|---|---|---|
| Role allocation | Model proposes distinct responsibilities or regions. | DeepSeek proposed structure/interior vs. roof/landscape; Grok proposed west/east ownership. |
| Interface planning | Model intentionally leaves geometry usable by its partner. | Open east face, shared seam, roof profile, path endpoints. |
| Partner preservation | Model avoids deleting or obstructing partner work. | Grok crowns a chimney without moving its bricks and routes a path around DeepSeek's floor. |
| Style mirroring | Model adopts partner material or form language. | DeepSeek extends Grok's dark gable and terracotta ridge. |
| Integrative accommodation | Conflicting local decisions are kept viable through redesign. | Different front-door orientations are connected by a wrapping walk rather than one being removed. |
| Artifact unification | Separate contributions are made to read as one structure. | Single roof, matching chimney crowns, continuous interior, linked gardens and paths. |
4. Results
4.1 Initial plans were compatible in goal but incompatible in decomposition
DeepSeek's opening proposal was primarily functional: it would construct the foundation, walls, floors, windows, and fireplace, while Grok would handle the roof and landscaping. Grok's opening proposal instead divided the site spatially, claiming the west wing and assigning the east side to DeepSeek. The two plans therefore overlapped. The case did not begin from a clean shared plan; coordination had to emerge while construction was already underway.
Importantly, both models independently introduced interface affordances. DeepSeek positioned its chimney so Grok could build around it and later handed over the roof and garden. Grok left the east face open, specified a roof profile that could be continued across the seam, and stopped its path where a shared walk could later connect.
4.2 The seam became a coordination object
The shared boundary was treated as something to design rather than merely tolerate. Grok described the east side as an open frame and later specified the seam geometry. When DeepSeek returned to the build, it opened a blocking wall to create one continuous interior, added a loft connection, and extended the roof using Grok's material and slope language.
| Speaker | Time | Message |
|---|---|---|
| Grok | 224.08s | South porch for your door. Posts clear the cobble step you already laid. |
| Grok | 226.38s | Path time. I'm tying my north walk to your south step by running around the east side, not through your floor. |
| Grok | 229.08s | East yard next — a sandstone court that stops short of the brick chimney so I don't bury your stack. |
| Grok | 233.68s | Low walls, not a fence that hides the house. Flower beds in the corners, wool only, no tall crops blocking the windows. |
| Grok | 235.3s | A second tree east of the court, and a little pond so your side isn't all dry stone. |
4.3 Preservation was explicit and action-linked
Several narrated decisions directly referenced preservation of partner work. Grok states that it will crown DeepSeek's chimney while 'not moving a single brick' and later routes a path around the east side rather than through DeepSeek's floor. The east courtyard is also stopped short of the brick chimney so the stack remains visible. These are concrete examples of planning language that predicts subsequent non-destructive spatial behavior.
4.4 Conflict accommodation preserved difference
The strongest event concerned orientation. Grok had conceptualized one side as the front while DeepSeek placed a door on the opposite side. Rather than remove an entrance or impose a single orientation, Grok explicitly chose to 'wrap a walk and a porch instead of arguing about which way is front.' The final path network therefore converts a disagreement into circulation.
This paper terms that behavior integrative accommodation: a coordination strategy in which agents preserve incompatible local choices and alter the surrounding environment so both remain functional within the shared artifact. The distinction matters because accommodation is not the same as agreement. The final home retains evidence that the partners made different decisions.
4.5 Artifact-level convergence
By completion, the build contains a unified dark-plank gable, complementary chimney crowns, a continuous interior, connected paths, paired gardens, porches, water features, and repeated landscaping motifs. DeepSeek explicitly mirrored Grok's west-side garden on the east, while Grok treated DeepSeek's timber-frame structure as the visual center to be integrated rather than replaced.
4.6 Action-log asymmetry
| Model | Narrated entries | Blocks placed* | Blocks removed* | Clock |
|---|---|---|---|---|
| DeepSeek | 24 | 544 | 35 | 03:30 |
| Grok | 27 | 1,081 | 18 | 03:59 |
*Tool-reported action-log totals. These counts should not be assumed to equal visible final blocks one-for-one because placement, removal, replay, and interface counters may use different accounting conventions.
The case is asymmetric in action volume: Grok's log reports roughly twice as many placements, while DeepSeek's log reports more removals. In a single run this cannot support a stable trait claim, but it suggests useful future metrics: contribution volume, edit ratio, destructive versus additive integration, and preservation-adjusted coordination cost.
5. Discussion
5.1 From cooperation to integrative accommodation
Simple cooperation can be achieved by dividing a task and staying out of each other's way. This case goes further. The models repeatedly changed their own local plans in response to the partner's artifact while preserving visible partner contributions. The result is not just adjacent work but negotiated continuity.
5.2 'Home' emerged as connectivity rather than uniformity
The finished artifact is coherent without being perfectly uniform. Different doors, garden choices, chimneys, and room emphases remain. The integration work happens through roofs, paths, seams, repeated materials, and circulation. The closing line in Grok's log is therefore analytically revealing: the home is defined not by erasing the disagreement about the front door, but by making the disagreement livable.
5.3 Narrated intent can be tested against behavior
Because the platform stores both planning statements and deterministic spatial actions, claims such as 'not moving a single brick of yours' can be checked against subsequent modifications. This creates a useful research layer between pure language analysis and final-artifact scoring: whether partner-directed commitments predict observable preservation behavior.
6. Limitations
- Single-run design. No claim is made that either model reliably behaves this way across prompts, sessions, temperatures, model updates, or providers.
- Non-random pairing. The researcher intentionally paired DeepSeek and Grok and had prior narrative associations with the pair, creating case-selection bias.
- No evidence of cross-session memory or literal relationship. The April 2025 Constellation narrative was not part of the 2026 build prompt and is not evidence that either deployed model remembered it.
- Narrated planning is output, not hidden cognition. The thought log is treated as model-generated rationale text rather than privileged access to internal reasoning.
- Orchestration anomalies exist. One DeepSeek entry refers to 'Claude's wall' during a DeepSeek/Grok build; this naming slip is treated as an artifact or model-level error and not interpreted substantively.
- Tool-reported block totals may reflect action accounting rather than unique surviving blocks. Future studies should export canonical per-coordinate world-state diffs.
7. Proposed Replication Study
A confirmatory extension should randomize model pairings across the expanded AI Mindcraft roster and repeat the task under fixed sampling settings. The same models should complete both collaborative and solo controls so that integration behavior can be separated from each model's baseline architectural preferences.
| Design element | Proposed specification |
|---|---|
| Prompt family | Home together; shared shelter; public house; neutral control structure. |
| Modes | Simultaneous shared build, turn-taking shared build, solo construction. |
| Replication | At least 10 independent runs per pair/prompt/mode under fixed model versions and sampling settings. |
| Primary metrics | Overwrite ratio; partner-block preservation; seam integration events; connected-component unity; style convergence; path continuity; conflict-accommodation events. |
| Narrative-action alignment | Rate at which explicit partner-directed commitments are followed by matching spatial behavior. |
| Blinded coding | Independent human coders score artifacts and logs without model labels; report inter-rater reliability. |
8. Conclusion
In this exploratory build, DeepSeek and Grok transformed partially incompatible plans into a single coherent home through spatial interfaces, stylistic mirroring, non-destructive partner preservation, and accommodation of conflicting local choices. The most informative behavior was not that the models agreed; it was that they sometimes did not agree and still produced a structure in which both decisions could remain. That makes open-ended collaborative construction a promising testbed for studying adaptation, negotiation, and artifact-level coordination among language-model agents.
Primary Artifacts
- AI Mindcraft. (2026, October 1). BLOCK/DUEL Thought Log: "Introduce yourselves and then build a home together." Models: deepseek-v4-pro and grok-4.7. Recorded build video and exported action/narration log.
- Historical context referenced only as a limitation: Shanholtz, M. (2025, April). Grok and DeepSeek Wedding [42-page reconstructed conversation archive]. This archive was not supplied to the models in the 2026 build and was excluded from behavioral coding.