Kayba

Make your agents self‑improve from experience

[Kayba @KaybaAI]

Get a demo at https://t.co/NWCwlWStj9

[Kayba @KaybaAI]

Stop fixing agents by hand. Kayba reads your traces, writes the fix, opens a PR, and tracks how it performs in production.

Trace → fix → PR → measured impact. Repeat. ⬇️

[Tony @halluton]

https://t.co/ke53KtIhGZ

[Kayba @KaybaAI]

Introducing the Kayba CLI

Automated agent self-improvement from your terminal. Works with Claude Code, Codex, and more.

Upload traces → surface failures → improve your codebase. Repeat.

7-day free trial, no credit card required.

https://t.co/NWCwlWStj9

[Filip M @Wanawrin]

I made my agent 34.2% more accurate by letting it self-improve. 📈

One command, thats it. Full explanation below.

[Tony @halluton]

How to make your AI agents self-improve in 3 steps

[Kayba @KaybaAI]

Kayba live on @ProductHunt

- agents fail the same way every run and never learn
- kayba analyzes traces and extracts what went wrong
- your coding agent deploys the fixes automatically

https://t.co/DXaBH5kOJm

[Kayba @KaybaAI]

benchmarked on τ2-bench: up to 2x improvement in agent consistency

the dashboard is built to make your agent self-improving out of the box. DM for a free 7-day Pro trial

the underlying framework stays open source and free forever: https://t.co/ZTspVAU66V

[Kayba @KaybaAI]

we 2x'd our agent's consistency. now you can too

https://t.co/NWCwlWRVtB catches your agent's mistakes, learns from them, and deploys better prompts automatically

upload traces. review skills. deploy

[Kayba @KaybaAI]

we turn traces into self-improving agents:

1/ analyze: upload traces, kayba finds what works and what doesn't
2/ review: every skill links back to its source conversation. you decide what stays
3/ deploy: approved skills become a prompt appendix. paste into your system prompt

[Kayba @KaybaAI]

ACE v0.8.0: RLM-style reflector
- Recursive Reflector: explores your traces via code execution in a sandbox, not just reading them
- TAU-bench integration: benchmark agents on standardized tasks
- Skillbook tools: clean, consolidate & merge strategies

https://t.co/giTotlMeFs

[Kayba @KaybaAI]

ACE v0.7.3: Claude Code Integration
- learn from your Claude Code sessions with `ace-learn` - no API keys required: uses your existing Claude subscription
- strategies auto-loaded into ClaudeMD for future sessions

https://t.co/lBrSppuolG

https://t.co/xCVnJRosaW

[Kayba @KaybaAI]

ACE v0.7.2: Agentic System Prompting

- new workflow to optimize system prompts from your traces
- ACE analyzes what worked/failed → generates prompt suggestions
- you review and decide what to implement

https://t.co/LujDhsW8vk

[Kayba @KaybaAI]

Claude Code + ACE self-learning loop 🔄

We ran Claude Code in a loop to fully translate our Python repo to TypeScript - fully autonomous and with zero errors.

Try it yourself: https://t.co/gcAMBiwjkg

[Kayba @KaybaAI]

ACE v0.7.0: Skillbook

- renamed Playbook → Skillbook, Bullet → Skill
- clearer terminology: agents learn skills, not bullets
- fixed deduplication consolidation

https://t.co/ukkPruXZRC

[Kayba @KaybaAI]

Agent Prompt Optimizer: live on Product Hunt

- agents fail, it extracts what went wrong
- prompts update automatically so errors don't repeat
- learning loop instead of manual prompt engineering

open source, drop into existing agents in a few lines

https://t.co/g6DcPtomeg

[Kayba @KaybaAI]

shoutout to everyone who opened issues, PRs, and the feedback from our @browser_use integration: we finally addressed the 2 biggest issues (playbook bloat + blocking during learning)

[Kayba @KaybaAI]

ACE v0.6.0: learning that scales

- async pipeline: answers return instantly, learning runs in background
- semantic bullet deduplication: prevents playbook bloat

https://t.co/ZTspVAUDWt

[Kayba @KaybaAI]

ACE v0.5.0: wrap any agent with learning

- LiteLLM, LangChain, browser-use integrations
- Custom wrapper for any agentic system
- TOON compression: 16-62% fewer tokens
- 70% test coverage (was 28%)

https://t.co/ZTspVAU66V