Workbench
ModStage
Run Codex, Grok, Gemini, Kimi, Qwen Code, OpenCode and Claude subagents side by side inside Claude Code, and watch, steer and approve every step.
Loading content…
Runtime · Agent Amplifier
Runtime Amplification Layer for AI Coding Agents — Claude Code Hooks That Don't Just Block, They Amplify
Agent Amplifier is the runtime amplification layer for AI coding agents — the missing half of the hook layer. Modern coding agents (Claude Code, Cursor, Copilot, LangGraph, CrewAI, AgentScope, LangChain) all expose lifecycle hooks. The community filled them with guardrails — products whose job is to BLOCK dangerous actions. That work is necessary, but it's only half the job. Hooks can also amplify: dynamically shape WHAT the agent does next based on prompt complexity, drift signals, convergence, and budget. Agent Amplifier fills that empty slot with five deterministic Python hooks — effort router (5-tier complexity classifier), goal anchor (anti-drift re-injection), convergence detector (LTI stability), persona escalation (audit pressure per iteration), and token budget controller. Zero extra LLM calls, zero network, fail-open, AGPL-3.0. Dogfooded across 1.71 billion tokens of real coding sessions before v1.0.0 shipped on 2026-05-13. 1,972 tests, 100% branch coverage on files touched, mypy --strict.
1,972Tests
100%Branch Coverage
7Host Adapters
1.71BTokens Dogfooded
Figures as reported by the project. Check them against the code and paper linked above.
See it run
Watch it work.
Install
Run it locally.
Install
pip install agent-amplifierWhat it does
How it works.
Effort Router (5-Tier Classifier)
Deterministic complexity classifier routes each user prompt to one of five effort tiers. Shapes phase budget + model tier suggestion without an extra LLM call.
Goal Anchor (Anti-Drift)
Re-injects the original goal every N tool calls to prevent multi-turn drift. Counters multi-turn drift at the hook layer.
Convergence Detection (LTI Stability)
Detects when agent output is substantially similar to the prior iteration. Stops the loop. Saves tokens.
Persona Escalation
Each iteration ramps audit pressure: senior reviewer → security → principal → distinguished AI-safety, plus custom personas. Catches what single-pass review misses.
Token Budget Controller
Cost-bounded amplification. Set a session budget; AA shapes phase prompts to stay within it without truncating mid-task.
7 Host Adapters
Claude Code, Cursor, GitHub Copilot, LangGraph, CrewAI, AgentScope, LangChain. One install command per host: agent-amp install <host>.
SLM Memory Provider
Built-in integration with SuperLocalMemory. recall() before each prompt, remember() at session end. Local-first, zero cloud.
Streamlit Dashboard
Live telemetry: token bar, classification distribution, persona escalation timeline. agent-amp dashboard to launch.
Use it for
Where it fits.
- Reduce token usage in long Claude Code sessions
- Stop multi-turn drift on complex refactors
- Auto-escalate audit pressure on critical PRs
- Cross-host amplification (one hook layer, 7 agents)
- Cost-bounded agent runs with hard token budgets
- Convergence-aware iteration (no infinite loops)
Changelog
Recent changes.
This repository has no tagged releases yet; these are its latest commits.