cognify.

Applied R&D · AI · Data · Analytics

Tessera: Artificial Collective Intelligence

← Back to logbook

Can a team of agents learn to collaborate?

Starting point

Tessera was born as a research and design hypothesis: we do not want to limit agents to fixed specialties or assemble a collection of “roles” that repeat the same response. We want to explore whether a set of polymath agents can, over time, build team capability: assertive communication, attention distribution, useful critique, common memory, and situational leadership.

The question guiding the study is whether, throughout the resolution of real-world problems and, eventually, human feedback, an artificial collective intelligence can emerge as a capability that is not merely the sum or a voting process of individual intelligences.

Initial hypothesis: with persistent memory, distributed information, disagreement protocols, and results evaluation, a collective of agents can improve its coordination in a verifiable way.

What was tested

The first version of Tessera was executed locally with Qwen via Ollama. Five instances of the same model were configured with different perspectives:

  • Atlas: framing, systems thinking, and success criteria.
  • Lumen: alternatives, weak signals, and information gaps.
  • Forge: experimentation, operationalization, and testing criteria.
  • Aegis: critique, risks, contradictions, and traceability.
  • Kernel: facilitation and integration, without authority to delete disagreements.

The cell worked in two rounds: initial positions and cross-objections. The system saved transcripts, contributions, and summaries of each mission.

First lesson: roles do not create a team

The test showed that assigning labels is not enough. Several instances of the same model, with the same context and similar biases, can produce a convincing but redundant conversation. In a career guidance mission, the collective proposed generic formulas without sufficient documentary evidence.

The discovery: "collaboration theater".

A multi-agent system can produce collaboration theater: multiple voices that aggregate text, but not knowledge.

Therefore, the criterion will not be the eloquence of the synthesis, but the traceability between evidence, hypothesis, decision, and subsequent result.

Working thesis

A collective displays collective intelligence when it achieves, under a comparable budget, results that an isolated agent or an aggregation of isolated agents does not obtain. Some possible signals:

  1. It builds a shared representation that no agent possessed at the start.
  2. It detects and corrects errors through substantiated objections.
  3. It distributes research and attention adaptively.
  4. It transmits relevant information without copying the entire context.
  5. It adjusts leadership, participants, and method according to the problem.
  6. It changes its future behavior based on observed results.

Tacit collective memory

Tessera seeks evidence regarding tacit and auditable collective memory. Is an agent or a collective of agents the same before and after a history of problems it has solved?

That memory could be expressed as:

  • which sources tend to be reliable for a certain problem;
  • which assumptions have failed before;
  • when to ask for evidence before proposing;
  • when to debate, when to research, and when to execute;
  • which type of agent or access is convenient to activate;
  • which historical disagreements remain open and under what conditions they would be resolved.

The memory should not become an opaque accumulation of text. Each entry should have an origin, date, confidence level, scope, evidence, subsequent result, and revision condition.

The six initial artifacts

1. Tessera Constitution

Executable rules to protect collective performance above individual showmanship. It includes distinguishing observation, hypothesis, proposal, objection, request, and decision; declaring uncertainty; and not inventing consensus or access to information.

2. Each agent's passport

A dynamic file for each participant: capabilities, depths, access, available memory, calibration, reliability history, and coordination style. The passport does not fix an eternal identity: it allows the role and authority to change according to the task and evidence.

3. Common knowledge registry

A compact registry of decisions, evidence, disagreements, results, and pointers to sources. It does not save complete transcripts as a substitute for memory. It prioritizes information that is recoverable, verifiable, and useful for subsequent missions.

4. Message protocol

All contributions must be classifiable as:

text
OBSERVATION · HYPOTHESIS · PROPOSAL · OBJECTION · REQUEST · DECISION

The objective is to reduce ambiguity, prevent a proposal from being disguised as a fact, and make visible what needs validation.

5. Cell formation policy

Rules to decide which participants intervene, what information they receive, who coordinates temporarily, and when an agent must abstain. Leadership is situational; it depends on the problem, evidence, access, and history of success, not on a permanent hierarchy.

A team is a team not because of linguistic choreography, but because of performing as a team that solves problems that isolated individuals could not, or would do so less efficiently.

6. Team evaluator

An evaluation mechanism that measures quality, cost, latency, diversity, traceability, subsequent corrections, and improvement against a single-agent baseline. The decisive metric is whether cooperation adds real value, not if it generates more conversation.

Next experiment

The next mission will seek to compare the performance of a single model solving the tasks, the same model having different roles, or different agents running on different models.

Proposed learning cycle

text
mission → hypothesis or decision → result observed by a person
→ postmortem → validated memory → change in future coordination

Without observed results, a team of agents accumulates narrative. With feedback from the world, it can build judgment.

Open questions

  • What minimum conditions make useful cooperation appear and not just mere aggregation?
  • What size and diversity of agents justify the cost of coordination?
  • How is information transmitted without saturating memory and budget?
  • How can the collective change with experience without becoming opaque, complacent, or a prisoner of its own errors?
  • What metrics demonstrate that the team outperforms an individual agent with the same total budget?

Status: initial exploration. The statements in this entry are design hypotheses and observations of the first prototype, not scientific conclusions.