The Agor operating method
Agents should earn
the right to act.
Build around a measurable job. Separate context from behavior. Simulate failure before customers discover it. Release with receipts.
00
The minimum viable promise
Agor is not trying to automate a department in one prompt. The first release is one agent, serving one job, with a stated success criterion and an honest boundary. Everything below exists to protect that promise.
Name the job
Define one trigger, one person, one job, and one observable outcome. The agent is not allowed to hide behind activity metrics.
Build the context
Separate journeys, tools, rules, workflows, knowledge, memory, glossary, and response language so each can be tested and changed deliberately.
Add independent judgment
Supervisors watch policy and safety while the customer-facing agent stays focused on the conversation and the job.
Simulate the hard cases
Generate a broad set of realistic conversations, then score job completion, policy compliance, handoff quality, and repeated reliability.
Show the reasoning trail
Trace tool calls, knowledge lookups, supervisor actions, latency, and the reason a release behaved the way it did.
Close the learning loop
Turn human resolutions into reviewable knowledge drafts, re-run simulations, and ship a versioned improvement.
Commercial alignment
Pay for a job well done.
Outcome pricing is more than a pricing page. It forces the product to define success, identify exceptions, and show the measurement clock. Agor's MVP exposes the outcome beside every agent so that alignment starts in the interface.
Release discipline
No invisible changes.
Every live version should preserve its model, instructions, knowledge, tools, simulation result, and monitor set. Roll forward with evidence. Roll back without archaeology.