All work
Alfred — a multi-agent orchestration framework
Alfred: a production framework that routes multi-step work across ~45 specialised AI agents, with review gates and clean hand-offs. (How it works kept deliberately high-level.)
Focus:
- Multi-agent
- Orchestration
Type:
Framework (Alfred)
Discipline:
AI application engineering
Tags:
- #Multi-agent
- #Orchestration
- #Reliability
- #Alfred
Note: client identities and full system architecture are withheld for confidentiality — enough is shared here to show the shape and substance of the work.
Problem
A single monolithic prompt can't reliably handle real, multi-step work that spans research, building, review and delivery. It drifts, loses context, and leaves no clean place to insert checks or human approval — so it can't be trusted to run on its own.
Approach
- Alfred decomposes work into ~45 specialised agents, each with a narrow, well-defined remit.
- A coordinator routes each job and runs agents in parallel where it can.
- Independent review agents gate the work between stages, so weak output is caught before it ships.
- Designed to run largely unattended. (The inner mechanics are kept private.)
Result
- Complex jobs that one prompt could never hold together now run end-to-end, largely unattended.
- Review gates catch low-quality intermediate output instead of letting it reach the final result.
- Adding a capability is a matter of adding an agent, not rewriting one giant prompt — the system scales by composition.
Architecture
