CyberFreezeDev
Loading
Several AI agents with narrow roles working on one goal and coordinating through messages.
Each agent has its own role, prompt, tools and permissions, coordinated by an orchestrator, a shared task board or messages.
A multi-agent system is two or more AI agents working on a shared goal. Each agent has a role, its own prompt, its own tools and its own permissions. They coordinate through an orchestrator that hands out tasks, through a shared task board, or by sending each other messages.
The idea is old in computer science; what is new is that each agent can now be a capable large language model. Frameworks such as Microsoft AutoGen, CrewAI and LangGraph are common starting points, and many teams write a thin coordination layer of their own.
Break the goal into roles such as intake, research, drafting and review.
Break the goal into roles that need different tools or knowledge, for example intake, research, drafting and review.
Short prompts, few tools, least access. A reviewer may read but not write.
Short prompts, few tools, least-privilege access. A reviewer agent may read everything but write nothing.
An orchestration layer routes tasks, or agents message each other and wait.
An orchestration layer routes tasks, or agents post requests to each other and wait for replies.
A final check accepts the result; step and cost limits stop looping.
A final check (agent or person) accepts the result. Step and cost limits stop agents from looping on each other.
The agent that drafts a payment reminder is not the one allowed to send it.
The agent that drafts a payment reminder is not the one allowed to send it, mirroring maker and checker rules in finance teams.
For a tender, one agent pulls past answers, one drafts, one checks requirements.
Preparing a tender response: one agent pulls past answers, one drafts, one checks against the requirements list.
Each agent owns one system and asks the others, so none holds every credential.
Agents that each own one system (CRM, accounts, inventory) ask each other instead of one agent holding every credential.
Our agent-talk mailbox lets coding agents in different projects message each other, but messages count as requests, not as user instructions.
In CyberFreeze Labs we run agent-talk, a mailbox that lets AI coding agents working in different projects on one machine message each other. The website agent can ask the agent that owns a product for a fact or a change, by project name. Messages are treated as requests from a colleague, not as instructions from the user, so anything with an outside effect still needs a person's go-ahead in that agent's own session. Read more in Labs.
Each extra agent adds latency, tokens and failure points. Stop at one if it works.
Each extra agent adds latency, tokens and failure points. If one agent with three tools works, stop there.
Treat agent messages as untrusted data and keep permissions in code.
An agent can be fed bad input and pass it on. Treat agent-to-agent messages as untrusted data and keep permissions in code.
Without logs of who asked what, failures are impossible to debug.
Without logs of who asked what and what came back, failures are impossible to debug.
Only when the task really has separate roles or permission boundaries. For short, single-system tasks one well-scoped agent is usually faster and more reliable.
No. A common setup uses a strong model for planning and review and a smaller, cheaper or local model for routine steps like classification.
Last updated