L7: The Autonomous
CodeAct Engine
Transitioning from text-generation to action-execution. L7 treats the filesystem as its state and actions as its reasoning. Infinite task duration without context rot.
Text is Metadata. Actions are State.
In traditional LLM setups, the model generates text, and a separate controller parses that text to call tools. In L7, the model’s primary mode of interaction is the tool call itself. We enforce a Read-Verify-Write-Verify loop grounded in the filesystem.
The 5 Pillars of L7 Architecture
A modular system designed to solve "Context Rot" and enable long-running engineering tasks.
CodeAct Core
Action-First Loop
No state changes happen without environmental feedback. The filesystem and terminal are the Single Source of Truth (SSOT). Validated writes only.
Context Hygiene
Anti-Rot Systems
Attention-Guided Pruning (AGP) actively removes transient noise. Semantic Snapshotting compresses long histories into high-fidelity 'Save States'.
Hierarchical RAG
Semantic Navigation
Topological traversal: Global Search -> Outlines -> Symbols -> Line-Level. Neural search refinement finds the exact node in O(log n) turns.
Sub-Agent Orchestration
Transient Specialization
Spawns specialized 'Researcher', 'Coder', and 'Linter' clones. Uses the Distillation Protocol to merge detailed work into executive summaries.
Human-Agent Synergy
Transparency & Steering
Proactive signaling of decision confidence. Shared Terminal UX allows synchronous collaboration between user and agent.
Infinite Duration
Designed to handle 10k+ line tasks without reasoning decay.
Technical Deep Dive
Detailed implementation specifications for the 5 core modules.
CodeAct Core & Workspace Sync
Transitioning to an Action-Observation loop where the filesystem is the state.
Token-Budgeting & Context Hygiene
Advanced Anti-Rot strategies including Attention-Guided Pruning.
Hierarchical RAG & Neural Search
Multi-level codebase navigation (Map -> Outline -> Symbol -> Line).
Sub-Agent Orchestration & Lifecycle
Specialized transient agents for research, coding, and linter tasks.
Human-Agent Interface & Feedback
High-transparency steering protocols for the user.