01
The decision comes before the agent
Define the business decision, authority, consequence, evidence, and owner before selecting an agent topology.
Engineering manifesto
Engineering principles for machine intelligence systems
These principles guide design and delivery when agents and swarms participate in consequential work.
EnterpriseGovernmentPartners
Manifesto
Principles
The principles are deliberately operational and testable.
02
Use the smallest topology that works
Every additional agent adds coordination cost, state, permissions, messages, failure modes, and evidence obligations.
03
Authority is separate from intelligence
A system may generate excellent recommendations without receiving authority to execute them.
04
Policy must be executable
Important rules belong at enforcement points with attributable decisions and observable outcomes.
05
Evaluate the path, not only the answer
Measure delegation, tools, memory, policy, cost, latency, retries, recovery, and review.
06
State and memory have owners
Every context store needs purpose, scope, provenance, retention, deletion, and access boundaries.
07
Every consequential action is reconstructable
Link the initiating decision, task graph, identities, versions, calls, policies, approvals, and outcomes.
08
Failure is a designed workflow
Timeout, denial, contradiction, stale state, provider failure, budget exhaustion, and shutdown are normal architecture cases.
09
Models and frameworks are replaceable
Durable capability lives in contracts, controls, fixtures, evidence, runbooks, and organizational judgment.
10
Human capability should compound
Machine intelligence should improve the organization’s understanding and decision quality, not quietly remove them.
11
Unknown is a valid evidence label
Uncertainty should narrow authority and trigger investigation—not be hidden by confident language.
12
A system is not complete until it can be stopped
Containment, rollback, revocation, and recovery are first-class acceptance criteria.
Start with a bounded decision
Use the manifesto as a design and review checklist.
A principle becomes useful when it changes an architecture, scope, release, or authority decision.