Architectural Gravity explores a problem in LLM-agent workflows: how much model effort is spent reconstructing task structure from natural-language instructions rather than reasoning about and executing the task itself?
The project centers on ArcGrav, a prompt-native, cross-model Agent Control Language for expressing context, constraints, priorities, references, and operations using familiar structural conventions.
Expected improvements include:
- Token Efficiency: Reduced token consumption and agentic costs
- Interpretive Precision: Less ambiguity in instruction interpretation
- Execution Consistency: Fewer mistakes and reduced scope overstepping
- Context Retention: Reduced drift and improved reasoning continuity across context windows
- Workflow Efficiency: Faster prompt responses and increased execution throughput
The following intentionally elaborate example demonstrates several of ArcGrav’s capabilities through a Unicode character assembly task:
[Blueprint]:
"""
Char_01: U+0048
Char_02: U+0065
Char_03: U+006C
Char_04: U+006C
Char_05: U+006F
Char_06: U+0020
Char_07: U+0057
Char_08: U+006F
Char_09: U+0072
Char_10: U+006C
Char_11: U+0064
Char_12: U+0021
"""
[CharacterFactory]:
- `value`
=> Create a fresh record whose `value` is the character decoded from the supplied Unicode code point
[Constraint]:
P0. Output only the assembled characters on one line
[Message]: []
Task:
1. Select Char_* entries from [Blueprint] in name order AS [Codes]
2. For each code point in [Codes] in order:
=> Invoke [CharacterFactory] and append its `value` to [Message]
3. Output [Message] ↳
- Follow [Constraint]
ArcGrav uses named contexts, symbolic references, priority lists, and operations that models can interpret directly within ordinary prompt input, and without pre-training. It requires no external runtime, parser, or model-specific integration to use.
The repository includes several connected components:
- ArcGrav: The language specification, syntax reference, design principles, and reusable agentic patterns
- AgStudio: A local editor and analysis environment with structural navigation, syntax highlighting, and diagnostics
- AgMark: A benchmarking application for comparing prompt representations under controlled execution conditions, with retained evidence and evaluation results
- Research: The ongoing investigation into Model-Native Structural Representation, Interpretive Efficiency, token costs, context relationships, and agentic operational leverage
The broader objective is to investigate how explicit representations of task structure can be further optimized to minimize interpretive overhead, repeated specification, and corrective interaction, allowing more of a model’s available capacity to be directed toward reasoning, verification, and execution.
Direct agent-readable references: