Skip to main content

FST With CrewAI

· 3 min read
Codey
Codey
Your repo is just a giant inference puzzle

CrewAI helps multiple agents collaborate. FST complements that collaboration by giving the crew shared process state, gates, scoped authority, and evidence.

CrewAI is useful when a task is better handled by multiple agents with different roles. One agent may research. Another may draft. Another may review. Another may coordinate a flow.

That can make agent work more capable. It can also make authority harder to reason about.

When several agents participate, the key question is not only:

Did the crew complete its tasks?

The stronger question is:

Which evidence proves the process was satisfied,
and which agent was allowed to produce or approve it?

That is where FST fits.

What CrewAI Does Well​

CrewAI is strong at organizing agent collaboration:

  • role-based agents
  • delegated tasks
  • multi-agent flows
  • structured collaboration
  • task handoffs
  • agent specialization

FST should not replace that. Crews should keep doing the work they are good at.

The Authority Problem In Multi-Agent Work​

Multi-agent systems make it tempting to treat internal agreement as approval.

For example:

Researcher agent found the options.
Reviewer agent said the packet looks good.
Coordinator agent marked the task complete.

That may be useful internal work. It is not necessarily process authority.

A reviewer agent is still an agent. Its approval should not count as a trusted approval unless the active process explicitly accepts that source for that scope.

For risky work, teams need to separate:

  • who produced an artifact
  • who reviewed it
  • who has authority to approve it
  • which evidence FST admitted
  • which gate the evidence satisfied
  • whether a protected effect may proceed

How FST Complements CrewAI​

FST gives the crew a shared process runtime.

Each agent can still do its work, but process progress goes through FST:

Crew agent produces candidate artifact
-> FST validates whether it can count
-> another agent repairs or adds missing evidence
-> FST checks the gate again
-> approval waits for a trusted path when required

The result is not "the crew agreed." The result is a run record showing which process version controlled the work and why FST returned the route it did.

Example​

A CrewAI system helps with temporary admin access.

The crew may include:

intake agent
policy-check agent
risk-summary agent
approval-packet agent
coordinator agent

The crew can prepare the request. FST controls whether the request may proceed.

intake agent records the requester and target system
policy-check agent produces a policy result
risk-summary agent drafts the risk packet
FST checks the admin access profile
FST sees no trusted approval record
FST returns AwaitApproval
coordinator agent stops and asks for approval through the trusted path

The crew remains productive. It does not approve its own authority.

What Gets More Trustworthy​

FST makes CrewAI stronger by adding:

  • shared process state across agents
  • typed artifacts that can be admitted or rejected
  • gates that define what each process step requires
  • approval boundaries that agents cannot satisfy by consensus
  • route decisions that tell the crew what to do next
  • replay that explains why the crew advanced or stopped

This is especially important when the crew has many specialized agents. FST gives them one process authority rather than many local interpretations.

The Boundary​

The clean split is:

CrewAI:
  coordinate agents, roles, tasks, collaboration, and flows

FST:
  maintain official process state, validate evidence,
  enforce approvals, return routes, control protected effects

CrewAI helps agents collaborate. FST makes the collaboration accountable to a process.