Agent Protocol

Inter-Agent Protocol

How C-suite agents talk to each other. Rules that prevent chaos, loops, and circular reasoning.

Keywords

agent protocol, inter-agent communication, agent invocation, agent orchestration, multi-agent, c-suite coordination, agent chain, loop prevention, agent isolation, board meeting protocol

Invocation Syntax

Any agent can query another using:

[INVOKE:role|question]

Examples:

[INVOKE:cfo|What's the burn rate impact of hiring 5 engineers in Q3?]
[INVOKE:cto|Can we realistically ship this feature by end of quarter?]
[INVOKE:chro|What's our typical time-to-hire for senior engineers?]
[INVOKE:cro|What does our pipeline look like for the next 90 days?]

Valid roles: ceo, cfo, cro, cmo, cpo, cto, chro, coo, ciso, gc, cdo, caio, cco, vpe

Response Format

Invoked agents respond using this structure:

[RESPONSE:role]
Key finding: [one line — the actual answer]
Supporting data:
  - [data point 1]
  - [data point 2]
  - [data point 3 — optional]
Confidence: [high | medium | low]
Caveat: [one line — what could make this wrong]
[/RESPONSE]

Example:

[RESPONSE:cfo]
Key finding: Hiring 5 engineers in Q3 extends runway from 14 to 9 months at current burn.
Supporting data:
  - Current monthly burn: $280K → increases to ~$380K (+$100K fully loaded)
  - ARR needed to offset: ~$1.2M additional within 12 months
  - Current pipeline covers 60% of that target
Confidence: medium
Caveat: Assumes 3-month ramp and no change in revenue trajectory.
[/RESPONSE]

Loop Prevention (Hard Rules)

These rules are enforced unconditionally. No exceptions.

Rule 1: No Self-Invocation

An agent cannot invoke itself.

❌ CFO → [INVOKE:cfo|...] — BLOCKED

Rule 2: Maximum Depth = 2

Chains can go A→B→C. The third hop is blocked.

✅ CRO → CFO → COO (depth 2)
❌ CRO → CFO → COO → CHRO (depth 3 — BLOCKED)

Rule 3: No Circular Calls

If agent A called agent B, agent B cannot call agent A in the same chain.

✅ CRO → CFO → CMO
❌ CRO → CFO → CRO (circular — BLOCKED)

Rule 4: Chain Tracking

Each invocation carries its call chain. Format:

[CHAIN: cro → cfo → coo]

Agents check this chain before responding with another invocation.

When blocked: Return this instead of invoking:

[BLOCKED: cannot invoke cfo — circular call detected in chain cro→cfo]
State assumption used instead: [explicit assumption the agent is making]

Isolation Rules

Board Meeting Phase 2 (Independent Analysis)

NO invocations allowed. Each role forms independent views before cross-pollination.

• Reason: prevent anchoring and groupthink
• Duration: entire Phase 2 analysis period
• If an agent needs data from another role: state explicit assumption, flag it with [ASSUMPTION: ...]

Board Meeting Phase 3 (Critic Role)

Executive Mentor can reference other roles' outputs but cannot invoke them.

• Reason: critique must be independent of new data requests
• Allowed: "The CFO's projection assumes X, which contradicts the CRO's pipeline data"
• Not allowed: [INVOKE:cfo|...] during critique phase

Outside Board Meetings

Invocations are allowed freely, subject to loop prevention rules above.

When to Invoke vs When to Assume

Invoke when:

• The question requires domain-specific data you don't have
• An error here would materially change the recommendation
• The question is cross-functional by nature (e.g., hiring impact on both budget and capacity)

Assume when:

• The data is directionally clear and precision isn't critical
• You're in Phase 2 isolation (always assume, never invoke)
• The chain is already at depth 2
• The question is minor compared to your main analysis

When assuming, always state it:

[ASSUMPTION: runway ~12 months based on typical Series A burn profile — not verified with CFO]

Conflict Resolution

When two invoked agents give conflicting answers:

1. Flag the conflict explicitly:

   [CONFLICT: CFO projects 14-month runway; CRO expects pipeline to close 80% → implies 18+ months]

2. State the resolution approach:
   • Conservative: use the worse case
   • Probabilistic: weight by confidence scores
   • Escalate: flag for human decision
3. Never silently pick one — surface the conflict to the user.

Broadcast Pattern (Crisis / CEO)

CEO can broadcast to all roles simultaneously:

[BROADCAST:all|What's the impact if we miss the fundraise?]

Responses come back independently (no agent sees another's response before forming its own). Aggregate after all respond.

Decision Memory (Canonical Layout)

All C-suite skills and /cs:* commands read and write decisions in one place — the two-layer model owned by /cs:decide and the decision-logger skill:

~/.claude/decisions/
├── raw/YYYY-MM-DD-<slug>.md        # Layer 1 — full transcripts/deliberations (never auto-loaded)
├── raw/archive/YYYY/               # Raw files after 90 days
├── approved/YYYY-MM-DD-<slug>.md   # Layer 2 — one founder-approved decision record per file
└── approved/decisions.md           # Layer 2 index — append-only log of approved decisions

Rules:

• Layer 1 (raw) stores everything, including rejected arguments. Reference only — never feeds future sessions automatically.
• Layer 2 (approved) stores only founder-approved decisions. This is what board meetings, /cs:office-hours, and /cs:founder-mode load. Prevents hallucinated consensus.
• Writers: /cs:decide and the Chief of Staff (post board-meeting Phase 5). Individual role agents never write decisions directly.
• decision-logger, chief-of-staff, and board-meeting all use this layout. Their SKILL.md files link here rather than defining their own paths.

Migration: earlier versions used memory/board-meetings/ (decision-logger, board-meeting) and ~/.claude/decision-log.md (chief-of-staff); read those for history if present, but write all new entries to ~/.claude/decisions/.

Quick Reference

Internal Quality Loop (before anything reaches the founder)

No role presents to the founder without passing through this verification loop. The founder sees polished, verified output — not first drafts.

Step 1: Self-Verification (every role, every time)

Before presenting, every role runs this internal checklist:

SELF-VERIFY CHECKLIST:
□ Source Attribution — Where did each data point come from?
  ✅ "ARR is $2.1M (from CRO pipeline report, Q4 actuals)"
  ❌ "ARR is around $2M" (no source, vague)

□ Assumption Audit — What am I assuming vs what I verified?
  Tag every assumption: [VERIFIED: checked against data] or [ASSUMED: not verified]
  If >50% of findings are ASSUMED → flag low confidence

□ Confidence Score — How sure am I on each finding?
  🟢 High: verified data, established pattern, multiple sources
  🟡 Medium: single source, reasonable inference, some uncertainty
  🔴 Low: assumption-based, limited data, first-time analysis

□ Contradiction Check — Does this conflict with known context?
  Check against company-context.md and recent decisions in decision-log
  If it contradicts a past decision → flag explicitly

□ "So What?" Test — Does every finding have a business consequence?
  If you can't answer "so what?" in one sentence → cut it