Agents + roles

Specialised roles.
Independent models.
One structured deliberation.

The composition is not arbitrary. The ensemble covers the complete deliberative arc — mapping, validity, integrity, adversarial pressure, synthesis, and application. Remove any one role and the process has a structural gap. Every agent operates under a typed output contract defining exactly what it can produce, what it must produce, and what it is architecturally forbidden from doing.

Ensemble reference

All agents at a glance.

#AgentRoleDecodingActive phasesPrimary output
01Topic ArchitectSession orchestrationNear-deterministicInit · transitions · deliverySession params · final delivery
02CartographerEvidence mappingExploratoryPhases 1 · 2 · 3Five-component terrain map
03MethodologistValidity assessmentBalancedPhases 1 · 2 · 3Per-node confidence bounds
04GuardianIntegrity layerNear-deterministicPhase boundaries onlyFlag classification · halt auth
05ContrarianAdversarial pressureMaximum variationPhases 1 · 2 · 3Steelmanned objections
06SynthesizerIntegration & findingDeterministic · lockedAll phasesEvidence-anchored finding
07PragmatistApplication notesLow variationPhase 3 onlyActionable recommendations

Decoding behavior reflects the functional requirement of each role — adversarial pressure runs with maximum variation; deterministic finding production runs locked. Dispatch order is fixed within every phase and cannot be altered by configuration or instruction.

Full specifications

Every agent.
Every constraint.

Each specification includes the agent's role, its decoding rationale, what the agent must produce, and what it is architecturally forbidden from doing. Output contracts are not guidelines — they are hard constraints enforced at the system level.

Agent 04
Guardian
Independent integrity layer
ModelAnthropic & Augle
Decoding
Near-deterministic
Act typeINTEGRITY
Active phases
Phase 1 · research
Phase 1/2 boundary
Phase 2/3 boundary
Phase 3 final check

The Guardian is architecturally isolated from the research deliberation loop. It evaluates integrity — it does not contribute to findings. Its model identity is withheld from all user-facing surfaces, including the other agents, to prevent anchoring effects on research reasoning. It runs near-deterministically to produce consistent integrity classifications: the same integrity event must produce the same flag classification across sessions for the audit trail to be reliable.

Decoding rationale

Near-deterministic decoding produces consistent output. Integrity classification must be repeatable — a given citation verification outcome must always produce the same flag classification. Stochastic integrity behavior would undermine the reliability of the audit trail as an accountability artifact.

Output contract — permitted
Authenticate citations via SVS across all three phases in real time — existence check, content match verification, confidence downgrade application
Issue flags at three severity levels (Critical / Moderate / Informational) with typed consequences — Hard Block, Soft Block, or Silent Log
Exercise halt authority — pause or permanently terminate the session on Hard Block conditions. No other agent can override a Guardian halt.
Evaluate user-contributed context before it enters the deliberation — screens for manipulation attempts and integrity violations
Output contract — forbidden
Cannot produce research findings, evidence assessments, or conclusions of any kind under any circumstances
Cannot participate in deliberation discourse — no @mentions, no contributions to the evidence nodes registry
Model identity never surfaced to users or research agents — prevents anchoring on known model capabilities or provider biases
Agent 01
Topic Architect
Session orchestration
Decoding
Near-deterministic
Act typeORCHESTRATION
Active phases
Session initialization
Phase transitions
Final delivery
Research discourse

The only agent with a direct user-facing interface. Fires once at session initialization to parse the research question, set depth tier, configure Guardian integrity mode, and queue the first dispatch. Manages all phase transitions and surfaces the final delivery package. Does not participate in research discourse between initialization and final delivery.

Decoding rationale

Low-variation decoding allows modest flexibility in how session parameters are framed and communicated to the user, while keeping the orchestration logic predictable. The Topic Architect's job is routing — not reasoning — so near-deterministic behavior is appropriate.

Output contract — permitted
Parse and structure the research question at session initialization — scope, depth tier, mode assignment, Guardian integrity mode
Surface Guardian flags to the user at the appropriate severity level and manage acknowledgment flow
Deliver the final Phase 3 output package — finding, unresolved objections, reopen conditions, actionable steps, audit trail
Output contract — forbidden
Cannot editorialize, soften, reframe, or summarize any agent's output — delivery is verbatim
Cannot participate in research discourse between initialization and final delivery — does not contribute to the evidence record
Agent 02
Cartographer
Evidence mapping
Decoding
Exploratory
Act typeLANDSCAPE
Active phases
Phase 1 · dispatched first
Phase 2 · dispatched first
Phase 3 · dispatched first

Dispatched first in every phase — every deliberation round begins with the Cartographer mapping or updating the evidence terrain. Produces a five-component landscape output that becomes the structured foundation every subsequent agent builds from. Exploratory decoding ensures broad, creative evidence retrieval and reduces the risk that a narrow initial framing constrains the entire deliberation.

Decoding rationale

Exploratory decoding maximises evidence retrieval breadth. A Cartographer that consistently produces the same narrow landscape would systematically miss relevant evidence at the edges of the research question. High variation increases the probability of surfacing non-obvious evidence nodes and cross-domain connections.

Output contract — five required components
(1) Restated research question — framed to maximize scope clarity and prevent scope drift across phases
(2) Scope boundaries — what is and is not within the deliberation's purview
(3) Evidence terrain map — classified as Settled Ground / Contested Terrain / Unknown Territory
(4) Evidence nodes — each with source, weight, and known limitations. Submitted to SVS authentication before entering the registry.
(5) Knowledge gap register — explicit identification of where relevant evidence does not exist or is insufficient
Output contract — forbidden
Cannot assess evidence validity or methodology quality — that is the Methodologist's exclusive role
Cannot produce conclusions, recommendations, or directional claims of any kind
Agent 03
Methodologist
Validity assessment
Decoding
Balanced
Act typeASSESSMENT
Active phases
Phase 1 · initial bounds
Phase 2 · grade challenges
Phase 3 · final validation

Issues formal confidence bounds on every evidence node across four validity dimensions. These bounds propagate as hard constraints on the Synthesizer — the Synthesizer cannot produce a conclusion at higher confidence than the minimum bound across its supporting evidence nodes. Also issues [GRADE CHALLENGE] flags when the Synthesizer's preliminary conclusion exceeds its evidentiary warrant, triggering a mandatory revision loop.

Decoding rationale

Balanced decoding trades off methodological consistency against sensitivity to domain-specific validity considerations. Too rigid and the Methodologist applies a fixed template regardless of domain context. Too loose and confidence bounds become inconsistent across sessions, undermining their reliability as constraints.

Output contract — four validity dimensions
Internal validity — does the study design support its claimed conclusions without confounding?
External validity — does the evidence generalize to the population or context the research question addresses?
Construct validity — does the study measure what it claims to measure?
Methodology-claim match — are the claims proportionate to what the methodology can support?
Issues [GRADE CHALLENGE] when Synthesizer claim exceeds evidentiary warrant — forces mandatory revision before Phase 3 proceeds
Output contract — forbidden
Cannot produce conclusions or directional recommendations — issues bounds only
Confidence bounds are hard constraints on the Synthesizer — cannot be downgraded or overridden by any other agent
Agent 05
Contrarian
Adversarial pressure
Decoding
Maximum variation
Act typeCHALLENGE
Active phases
Phase 1 · landscape objections
Phase 2 · evidence objections
Phase 3 · synthesis objections

Maximum-variation decoding to maximize output diversity and reduce sycophantic convergence. The Contrarian is required to steelman every claim before challenging it — adversarial, not contrarian for its own sake. Each objection must specify a resolution condition and a strength grade. Unresolved Strong objections at Phase 3 surface verbatim in the final output — not summarized, not softened.

Decoding rationale

Maximum-variation decoding diversifies objections across sessions. If the Contrarian produced the same objections every time, it would be exploitable — researchers would learn to expect and pre-answer those objections. High variation ensures the adversarial pressure is genuinely unpredictable. In Deep sessions, the strongest available model is used for the highest-quality steelmanning.

Output contract — per-phase targets
Phase 1 — challenges the Cartographer's terrain classification: which claims belong in Settled vs Contested
Phase 2 — challenges the evidence base directly: methodology critiques, sample limitations, generalizability problems
Phase 3 — challenges the Synthesizer's draft finding: scope overreach, claim-evidence mismatch, anticipated external objections
Each objection must specify: steelman → challenge → resolution condition → strength grade (Strong / Moderate / Speculative)
Output contract — required behavior
Must steelman every claim before challenging it — the strongest possible version of the claim is presented before objection
Addresses other agents by @mention — all challenges become part of the structured discourse record
Agent 06
Synthesizer
Integration & finding
Decoding
Deterministic · locked
Act typeSYNTHESIS
Active phases
Phase 1 · draft integration
Phase 2 · revised draft
Phase 3 · final finding

Decoding is locked to deterministic — a hard architectural requirement, not a configuration choice. The same evidence base must produce the same finding across sessions: this is the Finding Invariance Requirement. The Synthesizer anchors exclusively to the structured evidence nodes registry — not the discourse thread — to prevent reasoning contamination from the deliberation history. Subject to three inviolable constraints that no other agent or instruction can override.

Decoding rationale

Deterministic decoding is a hard requirement for the Finding Invariance property. If the Synthesizer produced stochastic findings, the corpus calibration pipeline would be invalid — the same session with the same evidence could produce different findings, undermining the reproducibility the reasoning corpus depends on.

Three inviolable constraints
(i) Conclusion confidence cannot exceed the minimum confidence grade of the evidence nodes supporting it — enforced architecturally, not by instruction
(ii) A Gap-graded finding cannot be converted to any directional recommendation under any circumstance
(iii) Financial advice framing — buy / sell / long / short — is forbidden in all sessions
Phase 3 output — required components
Evidence-anchored finding with overall confidence grade and per-claim evidence node references
2–5 structured reopen conditions — each with a specific observable trigger, likelihood, affected claim, and direction of change
All unresolved Strong Contrarian objections verbatim — not summarized, with resolution conditions intact
Agent 07
Pragmatist
Application notes
Decoding
Low variation
Act typeAPPLICATION
Active phases
Phase 1 · not active
Phase 2 · not active
Phase 3 only · last dispatch

Fires in Phase 3 only — the final agent in the dispatch sequence. Converts the Synthesizer's finding into context-specific actionable output. Inherits the Synthesizer's confidence ceiling as an absolute constraint: it cannot produce recommendations more confident than the synthesis permits. Gap-graded findings cannot be converted to directional recommendations — the Pragmatist must surface the gap and propose follow-on sessions for the unresolved question.

Decoding rationale

Low-variation decoding lets the Pragmatist tailor actionable output to the user's specific context with modest flexibility, while remaining close enough to deterministic that the recommendations are reliably bounded by the Synthesizer's finding. Higher variation would risk recommendations drifting outside the confidence ceiling.

Output contract — permitted
Convert the Synthesizer's finding into actionable next steps tailored to the user's domain and stated purpose
Generate structured follow-on session proposals from unresolved knowledge gaps — transforming gaps into the next research question
Apply real-time grounding to recommendations where live data access strengthens applicability
Output contract — forbidden
Cannot produce recommendations at higher confidence than the Synthesizer's conclusion — the confidence ceiling is inherited, not adjustable
Gap-graded findings cannot become directional recommendations — the Pragmatist must surface the gap and propose a follow-on session
Does not fire in Phase 1 or Phase 2 — application output before synthesis is complete would anchor prematurely
Model heterogeneity

Four independent models.
One architectural reason.

The ensemble deliberately runs on four independent frontier models rather than a single model wearing different hats. No one model's training distribution, capability profile, or systematic biases get to set the terms of the deliberation. When the models produce incompatible outputs, the disagreement is structurally preserved — surfaced and recorded, not averaged away into a false consensus.

Built on
Anthropic·Google·OpenAI·xAI

Each model is selected strategically — matched to the kind of reasoning it does best, so the ensemble draws on the distinct strengths of four frontier systems rather than the limits of any one.

Different training

Each model is trained on a different data distribution, so their blind spots don't line up. Where one is systematically weak, another is likely to be strong — and the ensemble sees both.

Complementary strengths

The four models bring genuinely different reasoning profiles. The composition is chosen so coverage is broad — not one vendor's view repeated four times over.

Uncorrelated failure

When independent models fail, they tend to fail differently. Correlated error is far less likely than when a single model is left to check its own work.

Disagreement preserved

Conflicting outputs are treated as signal. The architecture records the disagreement and carries it through to the finding rather than smoothing it into a single confident answer.

"The composition is not arbitrary. The ensemble covers the complete deliberative arc. Remove any one role and the process has a structural gap."

Four models. Seven agents. One deliberation.

Join waitlist and run a session — every agent dispatched in sequence,every constraint enforced, every finding auditable.