Skip to content
aiengineer .uk

Agent and workflow pipelines

Multi step processes where a model calls real tools against real systems, built so a failure halfway through does not leave your data in pieces.

The problem

Agent demos are easy. Agents that touch a production database, retry safely and can be replayed after an incident are a different discipline.

What the work involves

  • Explicit state machines rather than hoping the model remembers the plan
  • Idempotent tool calls, so a retry cannot double charge or double send
  • Human approval gates on anything irreversible
  • Full execution traces, replayable after the fact
  • Budget and step ceilings, so a loop cannot run up a bill overnight

Typical stack

MCP Temporal Laravel Queues Celery Redis Python TypeScript

Indicative rather than fixed. The right choice depends on what you already run, and inheriting a stack your team knows usually beats introducing a better one nobody can maintain.

Got a agent pipelines problem?

Technical detail welcome. The more concrete the problem, the more useful the first reply.