NCF Agent Interconnect
Documentation07. Consultation and human approval

07. Consultation and human approval

LLM-optimized documentation: llms.txt.

Learn how agents obtain information and human decisions while unrelated work continues.

  • Topic: Decisions and authority.
  • People in the example: 3 human decision-makers: product owner, architect and maintainer.
  • Agent chats in the example: 3 AI agents.
  • Example environment: Separate chats on one or more computers.

Browse all 13 examples · User guide · Setup

Prompt for your AI

Copy the complete prompt into your usual AI chat to get a detailed guide in your language. It explains the example below using the product documentation; you do not need your own project or an installed copy of NCF Agent Interconnect.

Write a guide using example 07 — Consultation and human approval.

Full product instructions:
https://github.com/no-colours-fabric/ncf-agent-interconnect/blob/main/llms.txt

Complete catalog:
https://github.com/no-colours-fabric/ncf-agent-interconnect/blob/main/docs/scenarios/README.md

Read before writing:
Read the complete product instructions above and the linked user guides needed
to explain this example accurately. Use the catalog to find other examples;
reading all 13 prompts is not a prerequisite. If a required source is
unavailable, identify it and request an accessible copy. Do not invent its
contents or present unsupported details as documented behavior.

Create documentation:
Write a detailed, self-contained guide in the language of our conversation,
unless I request another language. Keep the situation, participants and main
events below. Supply coherent fictional details and synthetic sample data
where needed. I do not need to provide a project, have NCF Agent Interconnect installed or match
the example's equipment. Do not replace the story with my real work or ask me
to choose again.

Explain the example directly. This selected example also satisfies the
example required in a general overview of NCF Agent Interconnect; do not add another or restart
selection. All dialogue, checks and outcomes belong to the illustration.
Produce the guide and its explanatory visuals only; do not connect agents,
send real messages, change files or settings, install software, run tests or
publish.

Example to explain:
Demonstrate task #T-1107: change an existing search API while preserving
compatibility. Use three human decision-makers: the product owner, architect
and maintainer. Three AI agents hold separate assignments: Alice investigates
requirements, Theo designs the contract and Leo implements the approved
change.

Distinguish consultation, which supplies information, from approval, which
authorizes an action. Alice clarifies contradictory requests; Theo asks the
human architect to choose between two designs; Leo requests the maintainer's
permission before changing the public API, with compatibility requirements
made explicit. Each request names the decision-maker, options, affected
artifact version and work that depends on the answer.

Record decisions in a shared project note. Missing answers leave dependent
actions pending; independent work continues. Silence is never approval.
Show the example's project owner
agreeing to a follow-up checkpoint. If reminders or deadlines must run
automatically, identify the separately configured external service; NCF Agent Interconnect
supplies no reminder scheduler and agents must not poll for approval.

Explain who made each decision and what the agents could do afterward. End
with the approved contract, remaining questions and the exact boundary of
authorized implementation.

Guide to produce:
Write a readable guide about the people's goal and how the work develops.
Open with the practical need, a compact participant table and only the setup
needed to understand the story. Link to the relevant setup guide; do not ask
the reader to perform setup.

Tell the story in order through four to six meaningful exchanges. Show short
messages as ordinary chat quotations with the sender and recipient named,
in the guide's language. A style example only; use the actual people, agents
and issue in this story:
**Sam → Leo:** "The date is missing from the screen. Can you check the field?"
**Leo → Sam:** "I found the mismatch. I'll correct it and ask you to check again."
Use these lines only as a formatting example. Write new messages about the
selected story, with its own participants, and translate them when needed.
After each exchange, explain in one or two sentences what the participants
learned, decided or did next. Make the value of direct messaging clear.

Keep messages readable. Do not show JSON, code fences, tool calls such as
agent_send, reply references, protocol fields, message IDs, raw attachment
objects or invented commit hashes as dialogue. If a technical detail explains
a user-visible problem, mention only that detail in brief prose outside the
quoted message. Summarize files or specifications by purpose rather than
pasting their contents. Use a simple example of the user's opening request,
in natural language; in English, "Connect as agent Name" is enough. Do not put
"AI" in words attributed to the user. Identify AI agents clearly in narration.

Give each working chat one main assignment. Example skills may come from
coordination, analysis, implementation and verification. Task, bug and issue
numbers such as #T-1101, #B-0321 and #I-0201 are story conventions, not
product requirements. Use the full product name NCF Agent Interconnect.
Distinguish its documented messaging from agent instructions and external
project tools. Message receipt is not task completion. Do not promise model
response times, universal idle-chat wake, transferred permissions or a task
scheduler. Use the current application support in llms.txt.

End with the illustrative result, what the people would check, human decisions
and remaining limits. Never present fictional checks as real evidence. Explain
the lesson without presenting this arrangement as a required workflow.
Keep the guide easy to scan. Do not add diagrams or illustration briefs unless
the reader specifically asks for one.

Finish with the catalog link above and invite the reader to choose another
example. Answer later specific questions directly, without repeating a whole
example or adding a catalog invitation. Adapt the language and detail when asked.