The runtime amplification layer for AI coding agents. Five deterministic Python hooks — effort routing, goal anchoring, convergence detection, persona escalation, token budgeting. No extra LLM calls, no network. Seven host adapters. AGPL-3.0.
Agent Amplifier ships SuperLocalMemory as its default memory provider. Each amplification turn recalls relevant context from SLM before the host agent runs and remembers key facts at session end. Two products, one composable runtime: SLM gives the agent persistent memory, Agent Amplifier shapes how the agent uses that memory at every turn.
Reference integration: examples/slm_provider.py in the Agent Amplifier repository.
No extra LLM calls. Fail-open — if any check errors, the host agent proceeds unmodified.
Deterministic complexity classifier routes each prompt to one of five effort tiers: minimal, low, medium, high, ultra. Shapes phase budget and model tier suggestion with no extra LLM call.
Re-injects the original goal every N tool calls to prevent multi-turn drift. The hook layer answer to the runaway-context problem.
Detects when agent output is substantially similar to the prior iteration. Stops the loop. Saves tokens and prevents oscillation.
Each iteration ramps audit pressure: senior engineer → adversarial reviewer → red-team auditor. Catches what single-pass review misses.
Set a session budget and Agent Amplifier shapes phase prompts to stay within it — without truncating mid-task.
One install command per host: agent-amp install <host>
Part of Qualixar · AGPL-3.0-or-later · Built by Varun Pratap Bhardwaj