What moves and what you rewrite
Two frameworks are exceptions. CrewAI builds its prompt from the
role, goal and backstory in your code, so a saved prompt reaches it only if you read it yourself. LlamaIndex and AutoGen are traced only: DecimalAI can’t hand either one a prompt or skills.
Before you switch
1
Put the prompt on DecimalAI
A prompt that still lives in your code won’t come with you. Open the agent’s Prompt page, paste the prompt your code uses today, choose Save draft, then Review & apply. From then on, the
fallback= in your code is only a backstop for when DecimalAI can’t be reached.2
Keep the agent's name
The new code passes the same name to
agent_config(). That name is what keeps the history together. If the name describes the old framework, such as langchain-react-bot, rename the agent before you switch, not after.3
Record a baseline
Score the old version now, so you have numbers to compare against. Attach server evaluators to the agent: they score every trace it sends, whichever framework sent it. If the old version runs on OpenAI Agents or the Claude Agent SDK, also run
decimalai.evaluate() on a set of cases and keep the report.4
Check how your skills reach the model
agent_config() delivers each skill’s text, its SKILL.md. Files a skill bundles, such as references/ and scripts/, reach the model only through a framework’s own skill loader, which reads them from .agents/skills/. If your skills depend on their files, pick a framework with a loader, and check that it can run scripts: the eight loaders lists what each one runs them with.Change the code
Withagent_config(), the switch changes the import, where the prompt and the load_skill tool go, and your tools. The agent’s name, its prompt and its skills stay as they are. Here is the same agent on LangChain and on OpenAI Agents:
- Before: LangChain
- After: OpenAI Agents
load_skill, and where to read how it takes your own tools:
Each framework’s page has the whole program, its install line, and how it continues a conversation.
Give each agent its skills one way. If the new framework has its own skill loader, point the loader at your skills or use
agent_config(), not both: doing both puts each skill in front of the model twice.Skills as files: change one line
If your skills are files in.agents/skills/, read by a framework’s own loader, the folder stays as it is. Only the line that points the loader at it changes: Setup at a glance has that line for each of the eight loaders. Claude Code, Cursor and other coding tools read the same folder, so the skills you try while you build are the ones your agent loads.
Check that nothing broke
- Skills still load. Open a few of the new version’s runs. Each run’s Recorded skills shows which skills were offered, delivered and loaded (what a run shows). A skill that loaded before and never loads now usually means the prompt or the
load_skilltool didn’t reach the model. - Scores held. Compare your server evaluators’ results on the agent’s Evals tab from before and after the switch. If the new version runs on OpenAI Agents or the Claude Agent SDK, run
evaluate()on the same cases and compare the two reports withcompare(). - No tool went missing. When the switch changes the tools, model or prompt that the new integration records, More → Code changes gets a new version that names what changed (what happens then). A tool you forgot to port shows up as a change to the tool list. Some integrations record only the tools a run called, so treat it as a hint, not a full inventory: what each one records.
Related
- Use skills in your framework: which way of adding a skill works on which framework
- Native skill loaders: skills as files, for the eight frameworks that load them
- Trust & Exit: what you keep if you leave DecimalAI itself