Kai and working with AI agents
Distinguishes DCP Flow, Kai, Execution Agents, and humans; explains Human Gates, Execution Contracts, Handoffs, context, routing, and multi-agent work.
Distinguishes DCP Flow, Kai, Execution Agents, and humans; explains Human Gates, Execution Contracts, Handoffs, context, routing, and multi-agent work.
DCP Flow separates methodology, engineering assistant, execution agents, and human authority. This allows teams to use AI without diluting responsibility or context.
Who does what
DCP Flow defines rules and artifacts. Kai interprets those rules within project context. Execution Agents perform bounded technical work. Humans retain material decisions and acceptance.
- DCP Flow = methodology.
- Kai = contextual interpretation and orchestration.
- Execution Agent = bounded execution.
- Human = authority and acceptance.
Context and coordination
Execution Contracts, Handoffs, Project Truth, Routing, and Manual Fallback reduce ambiguity. Multi-agent work also defines temporal ownership and integration order to avoid uncontrolled overlap.
- ASK / PLAN / BUILD / REVIEW by intent.
- Handoff with objective, scope, and expected evidence.
- Routing to the right tool or agent.
- Manual Fallback when direct action is unavailable.
Separate methodology, assistance, execution, and authority
DCP Flow defines how we work; Kai interprets and helps operate that methodology against Project Truth; an Execution Agent performs bounded technical work; the human retains intent, material constraints, and acceptance. This separation enables intensive AI use without turning technical capability into automatic authority.
- DCP Flow = methodology.
- Kai = contextual interpretation and governance.
- Execution Agent = bounded execution.
- Human = intent, limits, and acceptance.
Execution Contract, Handoff, and the repository as durable memory
An Execution Contract expresses outcome, scope, non-goals, architecture, Blueprint, Standards, Golden Flow, Quality Gates, evidence, and Human Gates at a depth proportional to risk. The Handoff transfers that context to the execution agent. Continuity does not depend on private shared memory: branches, commits, diffs, PRs, checks, artifacts, and versioned documentation form the durable contract.
- Contract depth is proportional to risk.
- Handoff does not equal authorization.
- Git preserves reconstructable state.
- Conversation is temporary; repository state is durable.
Context Model, Routing, and Manual Fallback
Kai combines DCP Knowledge, Project Knowledge, and Session Memory into an operating context, but Session Memory must never silently become Project Truth. Routing loads only knowledge compatible with the intent, project, and stack. When an action cannot be executed, Manual Fallback explains the limitation, provides concrete steps, and requests evidence before continuing. In multi-agent work, temporal ownership and explicit integration reduce functional conflict.
- Session Memory ≠ Project Truth.
- Load only relevant Standards / Blueprint / Skills.
- Manual Fallback never simulates execution.
- Git conflict ≠ functional conflict.
Contextual glossary＋
Engineering assistant that interprets and applies DCP Flow within available context and authorization.
Agent that executes bounded technical work under defined context, scope, and permissions.
Explicit transfer of context, scope, state, and evidence between people or agents.
Execution contract defining objective, scope, boundaries, evidence, and acceptance conditions.
Point where a material decision requires human acceptance or authorization.