Architecture
The first week is in the org, reading what already exists. Cut the process into steps, mark each one deterministic or adaptive, then choose the runtime. Agentforce where the permissions already live, a code agent where the exception lives, and a plain function where nothing needs judgement. The honest no: if the work is a software factory with a named delivery date, we are the wrong partner.
What it is
The runtime plane is where a step actually executes. Choosing it is a two-question test: is the step deterministic, and is a human waiting on it?
Why it matters
If you have only one seat to put work in, you will put the wrong work in it. These are the four we find.
Peers
Agent-to-agent delegation with the user’s identity carried across the wire. The gateway still sees every hop.
01
Declared skills
Each agent publishes what it can do and what scopes it needs. A registry, not a prompt that lists its friends.
02
A2A delegation
One agent asks another as a peer over a standard envelope, instead of a private wrapper per pair.
03
Identity travels
The user’s authority carries forward, so delegation cannot become privilege escalation.
04
Hop budgets
Traces, max hops, and timeouts stop cyclic delegation before it becomes an incident.
flowchart TD
U["User on a surface"] --> G["Gateway: identity, policy, audit"]
G --> B["Broker or registry: declared skills"]
B --> AF["Agentforce agent: org-native step"]
B --> CA["Code agent: SDK or LangGraph"]
AF <--> CA
CA --> T["MCP tools"]
AF --> R["Salesforce records"]In an org
The case step is org-native. The exception is not. Both are needed in the same process.
How Mindcat helps
We classify the steps before anyone picks a framework. That conversation is short and it saves a quarter.
Questions
Next
The brief
We will mark which steps are deterministic and which deserve a model.