Skip to content
Multigrid

Migrating pipelines and agents

Batch jobs, eval harnesses, guardrails, tool definitions and few-shot examples across a provider boundary.

Moving a single chat completion to another provider is an afternoon. Moving the machinery around it is not. A production pipeline carries a batch submitter that assumes one job lifecycle, an eval harness whose scores only mean something relative to the runs before it, a guardrails config written in one vendor’s safety vocabulary, a set of tool definitions in one schema dialect, and a body of prompts shaped by idioms the old model responded to. None of that is in the request body, and none of it moves by editing a base URL.

These pages work through that layer one piece at a time. The useful part of any mapping is the part that does not survive it: the status value with no counterpart, the parameter that means something subtly different, the default that differs, the capability one side has and the other does not. Each page names what is lost, what you rebuild by hand, and how you tell whether the rebuild worked before you cut traffic over.

Migrating Long-Running Batch Jobs Between Provider Batch APIs

How submission, polling and result retrieval differ between file-based and inline batch APIs, and how to rewrite a working batch script for a new target.

11 min read

Mapping Batch Job Status Values Between Providers

What each batch status value means, which ones are terminal, and why several have no counterpart on the other side at all.

10 min read

Migrating an Evaluation Harness to a New Provider

Why eval scores drop after a provider swap when the model is fine, and how to tell a real capability change from a harness artefact.

11 min read

Migrating a Guardrails Configuration Between Providers

How to rebuild a moderation ruleset when the target provider's safety categories are named differently, scoped differently, or do not exist as a configurable surface at all.

11 min read

What a Migration Does to Your Prompt Injection Defences

An argument that injection defences split into an architectural class that transfers intact and a behavioural class that does not, and that only the second needs re-testing.

10 min read

Migrating a Function-Calling Agent's Tool Definitions

A field-by-field conversion of tool definitions between provider schemas, including the result-message shape that changes with them.

11 min read

Why an Agent's Tool-Choice Behaviour Changes After a Provider Swap

Why an agent starts selecting the wrong tool when the definitions are byte-identical, and how to separate a translation bug from a genuine difference in tool-use behaviour.

10 min read

Mapping the "tool_choice" Parameter Between APIs

What each tool_choice value does, how the equivalents are spelled across API shapes, and which options exist on one side only.

10 min read

Migrating Response Format Enforcement ("json_object") Between Providers

What to do when response_format json_object is rejected on a migration target, and how to approximate the guarantee it gave you.

11 min read

What Changes in Refusal Behaviour After a Model Migration

How to detect, measure and reduce a refusal spike after a model swap, including the signals that distinguish a refusal from an ordinary short answer.

11 min read

Testing for Refusal Rate Changes Before Committing to a Migration

How to build a borderline prompt set from your own traffic and measure the refusal rate on both models, so a policy shift shows up in a test run rather than in support tickets.

10 min read

Conversation Summarization Breaks After a Provider Switch

Why a summarization prompt tuned to one model's compression habits produces summaries of the wrong length after a swap, and how to re-anchor it to a budget the model cannot reinterpret.

10 min read

Auditing Token Budget Assumptions Baked Into Chunking Logic

Why a chunker sized in characters or against a stale encoding starts producing oversized chunks after a model change, and how to find and re-derive every budget in the pipeline.

10 min read

Migrating Chunk Size Strategy When You Change Embedding Models

Why the chunk size that worked for one embedding model is not automatically right for another, and the arithmetic for re-deriving it and costing the re-index.

10 min read

Migrating an Agent Framework's Memory Store Format

How to convert stored agent conversation state between two frameworks' persistence schemas without breaking the tool-call invariants the target API will reject.

11 min read

What a Model Family Is, and Why Prompts Do Not Transfer Across One

Defines a model family by shared training lineage rather than shared vendor, and explains why that boundary predicts prompt portability better than the API you call.

9 min read

Auditing a Prompt Library for Provider-Specific Idioms

A review pass over stored prompts that flags the phrases, delimiters and hidden parameter dependencies that will not carry to another model family.

11 min read

Rewriting XML-Delimited Prompts for a Model That Does Not Favour XML

How to convert a prompt built on XML tags into a different delimiter convention without losing the boundary guarantees the tags were providing.

10 min read

Few-Shot Examples Stop Working After Switching Models

Why an example set tuned to one model's formatting habits degrades on another, and the specific adjustments — turn structure, count, balance, stylistic residue — that usually recover it.

10 min read

Other topics