MCP & AI workflow reliability
An AI workflow should
survive the handoff.
A good answer is only part of the job. An agent also needs to find the right tools, keep the right state and leave the next person with something they can use.
When a reconnect empties a tool catalog or a failed update looks successful, the work stalls even though the conversation looks fine.
Start with one
recurring failure.
Bring one workflow and the systems it crosses. It might be an MCP gateway that behaves differently after reconnecting, a source-linked answer that loses its evidence, or an accepted asset that never reaches its task record.
We first agree the process boundary and review permitted, redacted examples. We distinguish a reproducible defect from missing access, an unclear ownership decision or a limitation of the tools involved. That distinction determines whether a fix, a smaller pilot or a manual step makes sense.
Leave with a
useful decision.
The written scope can include a map of the handoff, a reproducible failure or clearly recorded evidence gap, proposed acceptance checks and an implementation recommendation. Where a pilot is useful, we agree the connections, review owner and failure handling before building it.
The checks should reflect the work: does reconnecting preserve a valid catalog? Does a result keep its supporting sources? Can a failed write be retried without duplicating the accepted output? These are examples to adapt to your process, rather than promises about an unexamined system.
Open to inspection
Inspect the engineering.
My independent MCP Context Forge contribution preserves the existing gateway catalog when discovery returns no tools, resources or prompts. Partial, nonempty discovery still removes stale discovered entries from absent catalog sections.
The merged PR shows the change and review. This is independent engineering evidence, not an IBM client engagement or endorsement.
Read the engineering notesStart with the process
Which handoff keeps getting stuck?
Send me the current workflow and the systems involved. A short description is enough to start; please leave out confidential data and credentials.
Email me about your workflowDelivery is remote, with English communication. Scope, fees, dates and allowed system operations are agreed in writing before paid work starts.