Agent Operating Model
Agents are not tools. They are accountable actors.
Dreamineering agents operate through shared language, declared capabilities, scoped authority, and receipts. Every action should be understandable before it runs and auditable after it completes.
The question is not whether an agent can act. The question is whether the system can explain who acted, under whose authority, toward what outcome, with what proof.
| Who? | did:web:mm.dreamineering.com |
| What jobs? | Five declared below |
| How call? | MCP endpoint — Phase B launch |
| What cost? | x402 micropayment USDC on Base — Phase B |
| How trust? | IntentTrace v1.1 receipt — public endpoint Phase B |
Five questions. One loop.
Operating Loop
Trust comes from the loop, not the label.
An agent becomes useful when intent can flow through language, capability, action, receipt, and consequence without losing accountability.
01
Intent
02
Language
03
Capability
04
Action
05
Receipt
06
Consequence
Four Surfaces
Read the system in the right order.
The capability registry is downstream of the operating model. First understand the language, compression, action design, and trust contract.
Language
DDL Nomenclature
Shared terms so humans and agents do not argue from different maps.
Compression
Dreamineering Symbols
Wire-format codes that let agents carry shared meaning at machine speed.
Action design
Actions and Consequences
A consequence lens for mapping output, action, changed state, and failure mode.
Trust contract
agent.json
Machine-readable identity, capability surface, governance, and call constraints.
Who This Is For
Start from the decision you need to make.
Owner / principal
Can I delegate without losing control?
Read the action cascade before authorising a bounded first delegation.
Developer / integrator
What can I safely call?
Inspect the manifest, verify capability status, and integrate only live contracts.
Operator / agent
What language do I use to coordinate?
Use DDL for human-readable terms and Symbols for compressed agent messages.
Declared Capabilities
Capabilities are contracts, not feature claims.
Five production-shaped capabilities are declared below. All status: planned — live MCP endpoints ship Phase B.
falsifiable-prediction
Convert any claim into a machine-verifiable bet: indicator + direction + threshold + check date.
Use when: Long-horizon decision needs a testable bet, not an opinion.
Input
{ claim: string, horizon_months: number }Output
{ indicator, direction, threshold, check_date, counter_case, maturity }flow-state-calibration
Given principal state (phase + pole dominance), return calibrated prompt for phase correction.
Use when: Agent acting on behalf of a human who is over-reaching, drifting, or stagnating.
Input
{ phase: sprint|recovery|reflection|calibration, dominant_pole: dream|engineer|reality }Output
{ prompt, timing, correction, pole_balance }pain-to-prd-classification
Extract pain signals from interview/research text. Score across 5 dimensions. Route to PRD/validate/park.
Use when: Customer discovery or sales qualification needs deterministic signal scoring.
Input
{ raw_text: string }Output
{ signals, scores, classification: create-prd|validate-demand|park, evidence_map }metric-definition
Transform a PRD outcome statement into a queryable metric: formula + threshold + unit + data source.
Use when: PRD or strategy output needs measurable proof, not prose.
Input
{ outcome_statement: string }Output
{ name, formula, threshold, unit, data_source, query }verifiable-intent-validation
Validate agent actions stayed within human-approved scope. Returns pass/fail + audit trail.
Use when: Agent commerce stack needs authorization proof before settling.
Input
{ intent_ref, actions_taken, delegation_chain, scope_constraints }Output
{ verdict: pass|fail, violations, audit_url, intent_trace }Machine-Readable Surfaces
Discover, evaluate, call.
Trust Model
Human in the loop. Always.
Kill Switch
Active — contact matt@dreamineering.com
Human-in-the-Loop
Required above spend threshold + first call from unknown caller
IntentTrace
v1.1 receipt on every tool call (public endpoint Phase B)
Governance
Two Funnels
Human Funnel
berley → hook → bait → fishball → platform
Humans discover via the berley trail — content scored for the right fish.
Same platform. Two funnels. One operating model.