Self-improving agent hierarchies - 4 plugins for quality feedback, auto-retry, skill routing, and prompt evolution

[bradysmith-dev]

I've built a set of 4 Paperclip plugins that add a self-improvement layer to agent hierarchies. Sharing here in case it's useful to others working on multi-agent setups.

Architecture

All 4 plugins are built with @paperclipai/plugin-sdk and communicate through Paperclip's JSON-RPC 2.0 protocol. Each runs as an isolated Node.js child process. They're event-driven and composable - use all four together or drop individual plugins into an existing setup.

The event flow:

issue.created  -> Skill Router assigns tools/agent type
issue.completed -> Performance Tracker records outcome, emits TASK_COMPLETED
                -> Self-Correction QA checks output (retry or escalate on failure)
                -> Prompt Evolver stores outcome for evaluation
                -> If degradation detected: Prompt Evolver proposes improved prompt

Plugins

Performance Tracker (@ai-corp/performance-tracker)
Subscribes to issue.completed events, logs success/failure outcomes, calculates rolling success rates, and emits TASK_COMPLETED with structured metadata. Detects degradation when success rate drops below a configurable threshold.

Self-Correction (@ai-corp/self-correction)
Listens for TASK_COMPLETED events from Performance Tracker. Runs a QA check on the output. If the check fails: retries up to a configurable limit, then escalates to a supervisor agent. Passes on success.

Skill Router (@ai-corp/skill-router)
Listens for issue.created. Analyzes the task description and assigns the appropriate tools and agent type before work begins. Prevents misrouting (e.g., a Worker agent receiving a task that needs architectural decisions).

Prompt Evolver (@ai-corp/prompt-evolver)
Subscribes to TASK_COMPLETED. Stores outcome data per agent. When Performance Tracker emits a degradation event, Prompt Evolver uses the accumulated outcome data to propose a rewritten prompt for that agent.

Test coverage and production status

• 56 tests passing across all 4 plugins
• Tested in production against a 3-agent hierarchy: CEO [Opus] -> CTO [Opus] -> Worker [Sonnet]
• Delegation chain verified end-to-end with all plugins active
• TypeScript throughout, Vitest test suite

What's for sale

The plugin pack is available on Gumroad: https://bradytech.gumroad.com/l/ai-companies

Includes the 4 plugins, a pre-wired company template (3-agent hierarchy with plugins configured), and a 10-chapter implementation guide.

$49 for the pack, $149 with a Q&A session and 30 days priority support.

I'm deaf so all support and discussion is text-based. Happy to answer technical questions here in this thread.