aislop Agent.

What it is

Most "AI code review" tools are LLMs reviewing code after a human writes it. The Agent is the inverse: an LLM that writes code, with the deterministic aislop engine acting as judge.

You point it at a path: /auth, packages/checkout, a single file, anything. It iterates:

1. 01 scan: runs aislop scan against the path, gets a score and a list of findings.
2. 02 propose: sends findings + file contents to the LLM, gets back a JSON diff (no shell access, no arbitrary code execution).
3. 03 apply: writes the proposed edits inside an isolated microVM (E2B sandbox).
4. 04 re-scan + gate: if the score regressed, revert the edit and try again. If it improved but is below threshold, loop. If above threshold, exit.
5. 05 PR: when the threshold is met, push to a fresh branch and open a pull request with the diff, cost report, and per-iteration trail.

Sample run

Real session output from a 6-file refactor:

# Spawn agent on /auth, target score 90, budget $1.00

[i] Cloning your-org/your-repo · sandbox up in 84ms
[i] Initial scan · aislop ci /auth → 78 / 100 · 9 findings

> Iteration 1 · LLM proposed 4 edits · $0.038
  - auth/middleware.ts:47  fix swallowed exception
  - auth/users.ts:12       remove narrative comments
  - auth/sessions.ts:8     drop unused import
  - auth/jwt.ts:91         narrow as any cast

[ok] Re-scan → 86 / 100 · 4 findings · accepted

> Iteration 2 · LLM proposed 3 edits · $0.041
[ok] Re-scan → 94 / 100 · 1 finding · accepted · threshold met

[i] Pushed branch aislop/agent/abc123
[i] Opened PR #274

────────────────────────────────────────
   94 / 100 Healthy · +16 from 78
   2 iterations · 6 files · 47 sec wall-clock
   Cost: $0.32 (LLM $0.30 + sandbox $0.02)

Why this is structurally different

Three properties no LLM-only tool has:

Cost model

Each session costs a few cents in compute, dominated by LLM tokens:

Per-session and per-org budgets are enforced server-side. The agent will not exceed your spend cap.

Pricing tiers

Sessions are bundled into hosted plans. See full pricing.

Trigger surfaces

• ★ Dashboard button: every project page has a "Spawn aislop agent" action. Live session view streams iteration deltas. (v1)
• ○ Slack slash command: /aislop fix /auth kicks a session, posts updates back to the channel. (v2)
• ○ PR mention: @scanaislop-bot fix the warnings in any PR comment spawns a session targeting the changed files. (v2)
• ○ Public-repo agent: open-source maintainers will be able to invite the agent on their repos via instant-scan URLs. (post-v1)

Safety guarantees

• Always opens a pull request, never direct-commits to any branch.
• Branch naming is deterministic: aislop/agent/<session-id>.
• The LLM proposes JSON edits; the host applies them. The LLM never has shell access inside the sandbox.
• Each session runs in a fresh microVM. Killed on completion. No persistent state.
• Per-session budget cap. Per-org daily budget cap. Per-org concurrency cap. None overrideable from the LLM.
• Full audit log: every spawn, every iteration, every cost event recorded.