LLSP

Software-engineering agent infrastructure

Independent authority and evidence for software-engineering agents.

LLSP separates model-generated proposals from repository execution authority. Proposed work crosses a typed boundary, local components decide whether mutation is authorized, and the controlled path records structured integrity evidence around the resulting action.

Model output is a proposal — not execution authority.
Read the non-confidential brief Contact the founder

The boundary

Control the software action, not the model’s wording.

Coding agents can inspect repositories, edit files, run commands, and delegate work. Sandboxes and runtime controls define where execution may occur. LLSP addresses a narrower repository question: what action was proposed, what was locally authorized, and what actually changed. The authority boundary is designed to remain separate from the model or agent that proposes the work.

Architecture at a glance

Proposal → typed boundary → authority → mutation → evidence

01
Agent / Model
Reasons about intent and proposes software work.
02
Typed Boundary
Converts proposed work into a structured contract before authority is considered.
03
Local Authority
Applies repository, path, permission, policy, and local-state checks.
04
Controlled Mutation
Authorized actions enter a governed mutation path with state checks and rollback behavior.
05
Integrity Evidence
Records structured mutation evidence, integrity records, and the resulting execution state.

Current status

Implemented core and current diligence are kept separate.

Implemented core

  • Typed separation between model-generated proposal and local execution authority
  • Strict contract validation before the governed mutation path
  • Repository/path boundary checks and local-state revalidation in the controlled path
  • Transactional mutation with checkpoint/rollback behavior
  • Structured integrity evidence around controlled repository mutations

Current diligence focus

  • External coding-agent adapters and direct-write containment
  • Delegated process, subagent, and MCP containment under the same authority boundary
  • Clean-environment reproducibility across target deployment environments
  • Signed or source-independent third-party verification
  • Broader adversarial and cross-agent benchmark evidence
Integration principle. Existing sandboxes, VMs, containers, and buyer-owned policy systems can remain the execution substrate. LLSP is intended to provide a repository-action authority and evidence boundary above or alongside that substrate, rather than replace it.

Technical & strategic inquiries

Start with the non-confidential architecture.

The public materials intentionally omit implementation-sensitive policy rules, verifier internals, bypass-corpus details, proprietary contract fields, and source code. Deeper technical material belongs in structured diligence after relevance and confidentiality are established.

Download brief founder@llsp.dev
Donghyun Lee Founder, LLSP
South Korea