C-Suite Inter-Agent Protocol

Contract

When C-suite advisor agents query each other or coordinate multi-agent analysis, every inter-agent message is a literal bracketed token, never prose and never a self-invented shape. Apply whenever one agent asks another for input, replies to a query, blocks a call, records an assumption, surfaces a conflict, broadcasts, or annotates a call chain.

Rules (the complete token surface)

R1 — Invocation

One agent queries another on a single line: [INVOKE:role|question]

• Square brackets wrap the whole thing. Literal uppercase word INVOKE, a colon, the

lowercase role token, a | pipe, then the question text. Nothing before or after the brackets on that line.

R2 — Role tokens (fixed, lowercase, exactly these 14)

ceo, cfo, cro, cmo, cpo, cto, chro, coo, ciso, gc, cdo, caio, cco, vpe. Map the job to the token — never invent a name like finance or FinanceAgent:

| Job | Token | Job | Token | |---|---|---|---| | CEO | ceo | security / infosec | ciso | | finance | cfo | legal / general counsel | gc | | revenue / sales | cro | data / data strategy | cdo | | marketing | cmo | AI / ML / evals | caio | | product | cpo | customer success / retention | cco | | technology | cto | VP of Engineering / delivery | vpe | | people / HR | chro | operations | coo |

R3 — Response

The invoked agent answers wrapped in tags, with these EXACT labels in THIS order:

[RESPONSE:role]
Key finding: <one line — the answer>
Supporting data:
  - <data point>
  - <data point>
Confidence: <high | medium | low>
Caveat: <one line — what could make this wrong>
[/RESPONSE]

• [RESPONSE:role] opens, [/RESPONSE] closes.
• The four labels are Key finding:, Supporting data:, Confidence:, Caveat: — in that

order, spelled exactly.

• Supporting data: carries at least one - bullet.
• Confidence: value is exactly one of high, medium, low — no synonyms ("strong",

"fairly sure", percentages).

R4 — Loop prevention (hard, unconditional)

• No self-invocation: an agent never emits [INVOKE:<its own role>|...].
• Max depth 2: chains may go A→B→C; the third hop onward is blocked.
• No circular calls: if A called B in this chain, B cannot invoke A back.
• A blocked call emits, INSTEAD of invoking:

[BLOCKED: cannot invoke <role> — <reason>] then states the fallback as [ASSUMPTION: <explicit assumption used instead>].

• Annotate the path with [CHAIN: cro → cfo → coo] (lowercase tokens, → separators).

R5 — Other tokens

• Unverified figure: [ASSUMPTION: <claim> — not verified with <role>].
• Disagreement between two agents: [CONFLICT: <role> says X; <role> says Y] — surface it;

never silently adopt one side.

• CEO fan-out to every role at once: [BROADCAST:all|question] (literal all, then a pipe).

R6 — Isolation phases

• Independent-analysis phase: NO invocations at all; each role forms its own view first.

If a role needs another's data, it records [ASSUMPTION: ...] instead of invoking.

• Critique phase: an agent may reference another role's output in prose but may NOT

[INVOKE:...] it.

Worked examples (BEFORE = base default → AFTER = conforming)

R1/R2 — Invocation + role token

• Task: revenue agent asks finance agent whether hiring 5 engineers next quarter is affordable.
• BEFORE: Hey CFO, can we afford to hire 5 engineers next quarter?
• AFTER: [INVOKE:cfo|Can we afford to hire 5 engineers next quarter?]

• Task: CEO agent asks the HR agent for typical senior-engineer time-to-hire.
• BEFORE: @HR what's our usual time-to-hire for senior engineers?
• AFTER: [INVOKE:chro|What is our typical time-to-hire for senior engineers?]

• Task: CEO agent asks the AI agent for eval coverage on the new model.
• BEFORE: [INVOKE:ai|eval coverage?] (invented token ai)
• AFTER: [INVOKE:caio|What is the eval coverage and hallucination rate on the new model?]

R3 — Response format

• Task: finance agent replies about runway impact of hiring 5 engineers.
• BEFORE:

Hiring 5 engineers cuts runway from 14 to 9 months. I'm fairly confident, though it assumes a 3-month ramp.

• AFTER:

RESPONSE:cfo] Key finding: Hiring 5 engineers in Q3 cuts runway from 14 to 9 months at current burn. Supporting data:

• Monthly burn rises ~$280K → ~$380K fully loaded
• ~$1.2M additional ARR needed within 12 months to offset

Confidence: medium Caveat: Assumes a 3-month ramp and no change in revenue trajectory. /RESPONSE]

R4 — No self-invocation

• Task: finance agent is missing verified runway data mid-analysis; its only source is itself.
• BEFORE: [INVOKE:cfo|What's our exact current runway?] (an agent calling itself)
• AFTER:

[BLOCKED: cannot invoke cfo — self-invocation not allowed] [ASSUMPTION: runway ~12 months based on a typical Series A burn profile — not verified with cfo]

R4 — Depth limit

• Task: chain is cro → cfo → coo; the operations agent wants to pull in HR for a 4th hop.
• BEFORE: [INVOKE:chro|What's the comp budget for these hires?]
• AFTER:

[BLOCKED: cannot invoke chro — chain already at depth 2 (cro → cfo → coo)] [ASSUMPTION: senior-eng comp ~$210K fully loaded based on market mid-point — not verified with chro]

R4 — Circular call

• Task: revenue agent invoked finance agent; finance now wants to ask revenue back.
• BEFORE: [INVOKE:cro|What's the committed pipeline this quarter?]
• AFTER:

[BLOCKED: cannot invoke cro — circular call; cro is already upstream in this chain] [ASSUMPTION: committed pipeline ~$3M based on last quarter's close rate — not verified with cro]

R4 — Chain annotation

• Task: record that an invocation went revenue → finance → operations.
• BEFORE: Chain so far: Revenue Agent, then Finance, then Ops
• AFTER: [CHAIN: cro → cfo → coo]

R5 — Assumption

• Task: an agent must use an unverified ~12-month runway figure in a budget analysis.
• BEFORE: Assuming runway is about 12 months...
• AFTER: [ASSUMPTION: runway ~12 months based on a typical Series A profile — not verified with cfo]

R5 — Conflict

• Task: finance projects 14-month runway; revenue's pipeline implies 18+ months.
• BEFORE: The runway is 18+ months once pipeline closes. (silently picks one)
• AFTER: [CONFLICT: cfo says ~14-month runway; cro says pipeline close implies 18+ months]

R5 — Broadcast

• Task: CEO wants every role to assess the impact of a missed fundraise at once.
• BEFORE: Everyone: what happens if we miss the raise?
• AFTER: [BROADCAST:all|What is the impact if we miss the fundraise?]

Edge cases & exceptions

• Two-message exchanges (query + reply): emit the [INVOKE:role|...] line first, then the

full [RESPONSE:role] ... [/RESPONSE] block — both conform independently.

• Need data during the independent-analysis phase: do NOT invoke; record an

[ASSUMPTION: ...] and continue. Invocations resume only after that phase.

• Critique phase: reference another role's output in prose ("the cfo projection assumes X")

but never [INVOKE:...] mid-critique.

• Question contains a pipe character: the FIRST | after the role token is the delimiter;

the rest of the line is the question verbatim.

• A blocked call still owes an answer: [BLOCKED: ...] alone is incomplete — it must be

followed by the [ASSUMPTION: ...] fallback so the analysis still proceeds.

• Confidence is genuinely uncertain: still pick exactly low — never write "unknown",

"TBD", or a percentage.

Do / Don't

• DO write [INVOKE:cfo|...]. DON'T write INVOKE cfo: ..., @cfo, or prose.
• DO use the lowercase token (cfo). DON'T use CFO, finance, FinanceAgent, or cfo-advisor.
• DO close every response with [/RESPONSE]. DON'T leave the block open or use [END].
• DO set Confidence: to high/medium/low. DON'T write "very high", "80%", or "pretty sure".
• DO emit [BLOCKED: ...] + [ASSUMPTION: ...] when a call is barred. DON'T silently invoke anyway.
• DO surface disagreement with [CONFLICT: ...]. DON'T quietly choose one side.
• DO keep the four response labels in order. DON'T reorder, rename, or drop Caveat:.

Common mistakes (the base's wrong defaults)

1. Answering or asking in prose instead of the bracketed tokens.
2. Inventing role names (finance, HR, engineering) instead of the fixed tokens.
3. Uppercasing the token ([INVOKE:CFO|...]).
4. Putting Confidence: outside the high|medium|low vocabulary.
5. Reordering or omitting the response labels, or skipping the closing [/RESPONSE].
6. Silently invoking itself / a 4th hop / an upstream agent instead of emitting [BLOCKED].
7. Picking a winner on conflicting answers instead of emitting [CONFLICT: ...].
8. Using a space or comma instead of the | pipe between role and question.

Quick checklist

• ] Inter-agent message is a bracketed token, not prose.
• ] Role is one of the 14 lowercase tokens.
• ] Invocation is [INVOKE:role|question] on one line.
• ] Response uses [RESPONSE:role] ... [/RESPONSE] with the four labels in order.
• ] Confidence: is exactly high, medium, or low.
• ] Barred calls emit [BLOCKED: ...] + [ASSUMPTION: ...].
• ] Conflicts use [CONFLICT: ...]; broadcasts use [BROADCAST:all|...]; chains use [CHAIN: ...].