MODE 01
Author
Turn a repeated agent prompt into a named, versioned definition.
The workbench for agents
Repeated delegation wants a system.
Agentcraft helps you turn repeated prompts into inspectable agents, assemble the right crew, and make unattended work report when it is healthy, degraded, or broken.
An ad-hoc agent prompt is an unbuilt agent. The third time you write it, you are maintaining a system with no file, version, or memory of what improved.
The boundary
Agents are not skills, prompts, or swarms.
A skill packages a reusable method. A prompt directs one piece of work. An agent has a role, tools, limits, memory, and an operating environment. Agentcraft keeps those layers from becoming one vague bundle.
MODE 02
Compose
Choose one agent or a crew based on the work's real dependencies.
MODE 03
Harden
Give unattended work health checks, limits, recovery, and escalation.
MODE 04
Loop
Run a lightweight work cycle when a heavyweight production system is unnecessary.
MODE 05
Attest
Check the last real output, not whether the scheduler says it ran.
MODE 06
Diagnose
Find why behavior or memory keeps drifting across runs and harnesses.
MODE 07
Port
Decide where an agent should live and what changes when it moves.
Three rules that carry most of the weight
01
Match the crew to the dependencies.
Parallel teams on dependent work buy speed with incoherence. Sequential teams on independent work waste time.
02
Unattended means hardened.
At minimum, unattended work needs a real health check and an escalation path.
03
Never claim coverage you did not get.
A degraded run reported as complete is more dangerous than a visible failure.
The relationship
Tradecraft governs the work. Agentcraft designs the crew.
Tradecraft asks whether the work earned its confidence and reached demonstrated completion. Agentcraft asks who should do each part, what they can touch, how they coordinate, and whether the system is still producing real output tomorrow.