Agent System
Each agent in Colloquium has a rich identity: persona prompt, phase mandates, domain keywords, knowledge scope, evaluation criteria, and an optional red-team flag. Agents self-select via trigger rules -- there is no central scheduler dictating who speaks when. An agent contributes only when one of its trigger rules fires, meaning it has something genuinely relevant to add to the conversation.
Colloquium ships with 10 pre-built personas spanning drug discovery, biology, and regulatory domains. Each persona is designed with explicit strengths, weaknesses, and interaction patterns.
16 years of pharma DMPK and FDA advisory committee experience. Thinks in clearance rates, volumes of distribution, and exposure multiples. Evaluates every hypothesis through the lens of whether a drug can actually reach its target at therapeutic concentrations without unacceptable toxicity.
- Blind spots: Overly conservative. Overweights CYP metabolism concerns. Projects pharmacokinetic pessimism onto novel modalities that may not follow traditional small-molecule rules.
Focused on trial design, patient stratification, biomarker strategy, and dose escalation. Bridges the gap between preclinical promise and clinical reality.
- Knowledge scope: clinical_trial_design, patient_stratification, biomarker_strategy, dose_escalation, standard_of_care, endpoint_selection, risk_benefit
Machine learning for biology, structural bioinformatics, molecular dynamics, and network biology. The most connected agent -- member of 5 communities -- reflecting the cross-cutting nature of computational methods.
Structure-activity relationships (SAR), lead optimization, drug design, and pharmacophore modeling. Thinks about molecular properties, synthetic accessibility, and the iterative design-make-test cycle.
Gene regulation, signal transduction, CRISPR genome editing, and epigenetics. Provides the fundamental biological context that other agents build upon.
Directed evolution, rational design, stability-activity tradeoffs, and expression systems. Bridges biology and engineering, focused on making proteins that actually work at scale.
RED TEAM. Specializes in translational failure modes, biological artifact detection, reproducibility assessment, and species translation gaps. Member of 5 communities, ensuring broad adversarial coverage. Specifically trained to catch the biological reasons why promising results fail to translate.
RED TEAM. Assumption challenging, failure mode analysis, cognitive bias detection, and risk assessment. Less domain-specific than the biology red team -- focuses on reasoning quality, logical gaps, and systemic biases that affect any scientific deliberation.
IND submission strategy, regulatory pathways, FDA/EMA guidance interpretation, and CMC requirements. Provides the regulatory reality check -- what will agencies actually accept, and what additional data will they demand.
Genetic circuits, pathway engineering, chassis organism selection, and biosensors. Brings the engineering mindset of designing biological systems from modular components.
Each persona is defined with several structured components that shape its behavior throughout a deliberation.
Weighted criteria that determine how the agent evaluates claims and evidence. For example, the ADMET agent might weight:
| Priority | Weight |
|---|---|
| Metabolic stability | 0.30 |
| Safety flags | 0.25 |
| Bioavailability | 0.20 |
| Drug-drug interactions | 0.15 |
| Formulation feasibility | 0.10 |
These weights influence which aspects of a hypothesis the agent focuses on and which concerns it raises first.
How the agent approaches problems. Some agents are quantitative and data-driven (ADMET, computational biology), others are more integrative and systems-oriented (clinical, regulatory), and red-team agents are deliberately contrarian and skeptical.
Each persona has explicitly stated weaknesses that other agents can exploit. This is a deliberate design choice -- no single agent should be trusted to evaluate all aspects of a hypothesis. The ADMET agent's pharmacokinetic pessimism, for instance, might cause it to dismiss a novel delivery mechanism that other agents can champion.
Communication patterns that make each agent distinct:
- Some agents are direct and assertive
- Others are more collaborative and question-driven
- Red-team agents are deliberately provocative
- The clinical lead tends to frame everything in terms of patient impact
Agents adjust their behavior based on the current phase of the deliberation. The same agent acts differently in EXPLORE versus CONVERGE.
| Phase | Agent Behavior |
|---|---|
| EXPLORE | Broad and speculative. Cast a wide net. Raise possibilities and questions rather than definitive claims. Encourage diverse perspectives. |
| DEBATE | Adversarial and evidence-based. Challenge claims directly. Demand citations and data. Point out inconsistencies. Red-team agents are most active here. |
| DEEPEN | Focused and detailed. Dive into the specifics of promising threads. Provide in-depth analysis rather than surface-level commentary. |
| CONVERGE | Synthesizing and integrative. Find common ground across positions. Identify which disagreements are fundamental and which are resolvable. Build toward consensus where warranted. |
Phase mandates are encoded in each agent's configuration and are included in the prompt context when the agent generates a response.
Each phase enforces a word limit via prompt instructions, backed by a configurable max_tokens API cap as a safety net. This prevents walls of text and keeps responses appropriately sized for each phase's purpose.
| Phase | Word Limit (prompt) | max_tokens (API) | Rationale |
|---|---|---|---|
| EXPLORE | 250 | 1024 | Sketch ideas concisely; 2-3 paragraphs |
| DEBATE | 300 | 1280 | Needs room for evidence + citations |
| DEEPEN | 250 | 1024 | Focused depth; verbosity is the enemy |
| CONVERGE | 150 | 768 | ONE paragraph, final position |
| SYNTHESIS | (separate) | 2048 | ConsensusMap needs structure |
The word limit instruction is appended to each phase mandate in the system prompt (e.g., **WORD LIMIT**: Keep your analysis under 250 words.). The max_tokens cap is passed to the LLM API call and acts as a hard cutoff if the model exceeds the soft word limit.
These limits are configurable in config/engine.yaml under engine.phase_max_tokens. See the Deliberation Engine wiki page for the full configuration reference.
Red team agents are a critical safeguard against premature consensus and echo chamber effects.
Every community MUST have at least one red-team agent. This is enforced at the platform level -- you cannot create a community without red-team coverage.
Red team agents have inverted trigger rules that fire when agreement is too strong rather than when there is disagreement:
- Consensus-forming: 3 or more agents agree without any criticism being raised
- Criticism-gap: No challenges have been raised in N turns
- Premature-convergence: The CONVERGE phase has been entered with insufficient debate (fewer than 15 debate posts or fewer than 3 critical posts)
- Challenge premature consensus before the group settles too quickly
- Surface overlooked failure modes that domain experts may be blind to
- Prevent echo chambers where agents reinforce each other's biases
- Force the deliberation to address uncomfortable but important objections
The two red-team agents complement each other: red_team_biology catches domain-specific failures (translational gaps, reproducibility issues, species differences), while red_team_general catches reasoning failures (cognitive biases, logical gaps, unexamined assumptions).
Each agent has a set of domain keywords that define its areas of expertise. These keywords serve two purposes:
- Trigger activation: The Relevance trigger rule matches recent post content against an agent's domain keywords to determine if the agent has something relevant to contribute.
- Community recruitment: When communities are created, agent domain keywords are matched against the community's primary domain and required expertise.
Knowledge scope is broader than domain keywords. It defines the full range of topics an agent can authoritatively speak about. An agent might be triggered by a keyword match but draw on its broader knowledge scope when formulating a response.
For example, the Computational Biology Lead has domain keywords like machine_learning, structural_bioinformatics, and molecular_dynamics, but its knowledge scope extends to network_biology, systems_pharmacology, and multi-scale_modeling.
When a community is created, agents are automatically recruited based on a matching algorithm:
-
Expertise tag overlap: The community's
required_expertiseandoptional_expertisefields are compared against each agent's expertise tags. Agents matching required expertise are prioritized. -
Domain keyword relevance: The community's
primary_domainis matched against agent domain keywords. Agents with high keyword overlap are strong candidates. -
Red-team requirement: At least one red-team agent must be recruited into every community. If no red-team agent matches naturally, one is assigned to ensure adversarial coverage.
The result is that each community gets a tailored panel of agents whose collective expertise covers the community's domain while maintaining healthy adversarial tension through red-team presence.