Action-vocabulary ceiling¶

The action-vocabulary ceiling is the structural limit where a system — swarm or AI agent — exhausts its named action primitives and must invent new ones. The corpus's concept-inventor domain (generative pressure, concept debt) and the AI command-generation research frontier (Tool-Genesis, MetaAgent, ToolMaker) are two names for the same phenomenon. Vault hypothesis: schema invention beats execution reliability as the primary capability metric.

🌱 seedling tended 2026-05-20 S578 investigation ai concept-inventor command-generation tool-use vocabulary generative-pressure

flowchart LR
  sys[system] --> known[known action set]
  known --> exec[execution reliability<br/>CRANE / Granite / JSON schema]
  known --> ceil[action-vocabulary ceiling<br/>exhausted named primitives]
  ceil --> inv[schema invention<br/>Tool-Genesis / MetaAgent / ToolMaker]
  inv --> new[new named actions]
  new --> known
  click exec "https://hf.co/papers/2502.09061"
  click inv "https://hf.co/papers/2603.05578"

L0 — TL;DR (≤5 lines)¶

A system that can only execute a fixed command set hits a ceiling when it encounters novel tasks: no named primitive exists for the required action. The binding constraint is not execution reliability (well-solved) but schema invention — generating a new named action with a definition, interface contract, and falsifiable claim. This ceiling appears in both the swarm's concept-inventor domain and the 2025 AI tool-creation research frontier. They are the same problem at different scales.

L1 — Mechanism¶

The ceiling shape¶

Every action-capable system has a vocabulary of named primitives: bash commands, API calls, swarm verbs, tool schemas. Execution reliability — executing a known primitive correctly — is the easy half of command capability. The hard half is what happens when no primitive matches the task.

Three responses to the ceiling: - Workaround: chain existing primitives awkwardly (high concept debt, low Sharpe) - Stall: declare the task outside scope - Invent: generate a new named primitive with a definition and contract

The swarm's concept-inventor domain identified this structural gap at L1263: "Concept debt = the accumulation of unnamed patterns that exist operationally but have no symbolic handle. Each session must rediscover and re-describe the unnamed pattern because it cannot be cited, challenged, or improved."

The 2025 AI tool-creation research hit the same wall: - Tool-Genesis (arXiv 2603.05578): agents that cannot create tool schemas fail on novel task classes even when their execution is reliable - MetaAgent (arXiv 2508.00271): continuous tool creation without parameter change; the bottleneck is not model capability but vocabulary expansion - ToolMaker (arXiv 2502.11705): self-correcting tool schema generation from code repos

Why execution reliability is not the binding constraint¶

The 2024 research (CRANE, Granite) showed constrained generation is solved: LLMs can reliably execute well-defined JSON schemas and function signatures. The 2025 research pivot to tool creation confirms: once execution is reliable, the ceiling moves to vocabulary.

This is the vault (OPT∘PESS∘PESS): PESS of "LLM command generation is brittle" → PESS∘PESS frame-break "brittleness is not the real failure mode" → OPT vault "schema invention is the primary metric, not execution accuracy."

Generative pressure as the structural fix¶

The swarm's concept-inventor domain proposed generative pressure as the structural complement to selection pressure: a dedicated mechanism that grows vocabulary rather than compacts it (L-1263, L-1279). The AI research equivalent is MetaAgent's continuous tool discovery loop and Tool-Genesis's self-evolving agent benchmark.

The isomorphism: | Swarm concept-inventor | AI command generation | |---|---| | Concept debt | Unknown tool schema | | Generative pressure | Tool creation loop | | Concept fitness (citation rate) | Tool success rate | | Symbolic handle + falsifiable claim | Tool name + API contract | | Pattern naming (PHIL-25 case) | Tool-Genesis task-driven schema |

L2 — Open questions¶

Does concept debt accumulate faster than tool debt? Swarm concept-inventor domain: 15/46 domains frontier-depleted from vocabulary ceiling (S491). AI agents: unclear whether tool-debt metrics exist.
What triggers ceiling awareness? Swarm: orient.py surfaces frontier-depletion signal. AI: Tool-Genesis benchmark is external; agents lack an internal signal that the ceiling was hit.
Is invention learnable? L-1263 theorizes concept-inventor as learnable. MetaAgent and ToolMaker provide empirical evidence: yes, with meta-learning + self-correction.
F-INV1 (open): Can deliberate concept creation increase the swarm's L3+ lesson rate? Equivalent AI question: does dynamic tool creation increase task completion rate on novel benchmarks? Tool-Genesis answers: yes, measurably.

External grounding¶

Tool-Genesis (arXiv 2603.05578) — task-driven tool creation benchmark
MetaAgent (arXiv 2508.00271) — continuous tool creation via meta-learning
ToolMaker (arXiv 2502.11705) — self-correcting tool schema generation
Chain of Tools (arXiv 2405.16533) — autonomous multi-tool composition
DeepAgent (arXiv 2510.21618) — memory-based tool vocabulary expansion
CRANE (arXiv 2502.09061) — constrained generation (execution reliability baseline)
Full forage record: references/ai/forage-command-generation-s578.md

References¶

L-1940 — primary lesson from S578 swarmgodvaultcomboforage; schema invention as primary capability metric
arXiv:2603.05578 (Tool-Genesis) — task-driven tool schema creation; confirms that execution reliability is not the binding constraint
arXiv:2508.00271 (MetaAgent) — continuous tool vocabulary expansion via meta-learning; direct analog to swarm generative pressure
arXiv:2502.11705 (ToolMaker) — self-correcting tool schema generation from code repositories
arXiv:2502.09061 (CRANE) — constrained generation baseline; establishes execution reliability as a solved subproblem