Coform
AM

Intelligence architecture · Internal

The invisible facilitatorthat powers Coform

Coform is not a chatbot. It is a multi-agent system that helps founders understand, align, discuss, decide, commit, and revisit — without ever standing between them.

Principle

Invisible

Users experience intelligence — not an interface to AI.

Principle

Specialized

Seven agents with one responsibility each — not one model doing everything.

Principle

Explainable

Every recommendation carries its reasoning. No black box.

Alignment pipeline

How intelligence moves through a session

  1. Context

    Context agent · Understands the organization

  2. Journey

    Question agent · Builds adaptive reflection journeys

  3. Adaptive questions

    Question agent · Builds adaptive reflection journeys

  4. Reflection

    Human moment · no AI intervention

  5. Insight analysis

    Insight agent · Analyzes completed reflections

  6. Discussion priorities

    Insight agent · Analyzes completed reflections

  7. Workshop

    Workshop agent · Prepares and facilitates conversation

  8. Agreements

    Agreement agent · Converts discussion into governance

  9. Memory

    Memory agent · Maintains organizational memory

  10. Future recommendations

    Recommendation agent · Continuously improves alignment

Seven specialized agents

One responsibility each

No single model does everything. Each agent has narrow scope, clear inputs, and outputs the next stage can trust.

Agent 01

Context

Onboarding · Session builder

Reads the outside world of the company so every downstream agent starts with shared footing — not a blank prompt.

Inputs

  • Website
  • Industry
  • Stage
  • Team size
  • Business model

Outputs

  • Company summary
  • Suggested focus areas
  • Industry context
  • Potential discussion topics

Boundaries

  • Never infers personality or diagnoses founders.
  • Public and org-provided signals only.
Future shape · Retrieval + summarization pipeline · MCP tool

Agent 02

Question

Reflection engine · Session builder

Shapes the questions each founder sees — ordering, difficulty, and follow-ups adapt to stage, journey, and history.

Inputs

  • Organization context
  • Journey type
  • Selected modules
  • Prior workshops
  • Prior disagreements

Outputs

  • Question set
  • Follow-up questions
  • Scenario questions
  • Ordering
  • Estimated duration
  • Question difficulty

Boundaries

  • Never quotes another participant's answer.
  • Difficulty adapts, but the question library is versioned and auditable.
Future shape · Prompt chain · workflow · question graph

Agent 03

Insight

Analyzing · Insights · Review

Synthesizes what the team said into shared foundations, quiet tensions, and the conversations worth having next.

Inputs

  • All participant responses
  • Confidence levels
  • Written context

Outputs

  • Shared foundations
  • Different perspectives
  • Needs agreement
  • Emerging themes
  • Hidden assumptions
  • Conversation starters

Boundaries

  • Never attributes a quote to a person.
  • Always synthesizes — never takes sides or ranks participants.
Future shape · Vector retrieval + clustering + LLM synthesis

Agent 04

Workshop

Review · Discussion · Workshop

Translates insights into an agenda a human facilitator can actually run — with time, order, and prompts.

Inputs

  • Insight set
  • Bookmarked tensions
  • Team availability

Outputs

  • Agenda
  • Facilitator prompts
  • Discussion ordering
  • Clarifying questions
  • Decision suggestions
  • Parking lot items

Boundaries

  • Suggests structure — humans choose what to discuss.
  • Never facilitates in-room in real time.
Future shape · Workflow · background job

Agent 05

Agreement

Constitution · Discussion

Turns decisions into durable language — with authorship, review dates, and the assumptions they rest on.

Inputs

  • Workshop decisions
  • Participants present
  • Referenced insights

Outputs

  • Founder Constitution entries
  • Decision wording
  • Agreement summaries
  • Review schedules
  • Ownership assignments

Boundaries

  • Never authors agreements alone — always drafts for human confirmation.
  • Every entry links back to the workshop that produced it.
Future shape · Prompt chain · MCP tool · document generator

Agent 06

Memory

Organization workspace · Insights

Holds the graph of people, sessions, agreements and topics — so alignment can be compared across quarters, not just captured in one.

Inputs

  • Agreement history
  • Workshop history
  • Session outcomes
  • Strategic pivots

Outputs

  • Alignment drift over time
  • Decision evolution
  • Topic timelines
  • Governance history

Boundaries

  • Compares — never rewrites — past answers.
  • Never surfaces private reflections outside their original scope.
Future shape · Graph store · vector index · background job

Agent 07

Recommendation

Dashboard · Organization workspace

Watches the graph over time and quietly suggests when a topic is due for a revisit, a review, or a new session.

Inputs

  • Memory graph
  • Confidence gaps
  • Time since last review
  • Recent pivots

Outputs

  • Quarterly review nudges
  • Revisit fundraising suggestion
  • Hiring alignment prompts
  • Outdated agreement flags
  • Shifting priority alerts

Boundaries

  • Suggests — never schedules or acts autonomously.
  • Every recommendation carries an explanation.
Future shape · Scheduled background job · rules + LLM ranker

System topology

Information flow between agents

Every agent reads from and writes to shared organizational memory. Forward hand-offs run around the ring; the graph runs through the center.

MemoryGraphContextAgent 01QuestionAgent 02InsightAgent 03WorkshopAgent 04AgreementAgent 05RecommendationAgent 07

Insight taxonomy

Eight kinds of insight

  • Shared beliefs

    The quiet foundation the team already agrees on.

  • Hidden assumptions

    What everyone thinks is settled — but never was said.

  • Different expectations

    Where the team is out of sync without knowing it.

  • Confidence gaps

    Answers given with low conviction, worth revisiting.

  • Missing agreements

    Decisions that have been made in practice but never in words.

  • Emerging themes

    Patterns across responses that hint at something structural.

  • Future risks

    Places where today's alignment could quietly break tomorrow.

  • Suggested conversations

    What the workshop should probably start with.

Insight schema

Every insight carries the same fields

  1. 01Summary
  2. 02Evidence
  3. 03Why it matters
  4. 04Suggested discussion
  5. 05Potential agreement
  6. 06Review recommendation

Decision intelligence

Every agreement answers seven questions

  1. 01Why was this decision made?
  2. 02Who agreed?
  3. 03When?
  4. 04What assumptions supported it?
  5. 05When should it be reviewed?
  6. 06What workshop produced it?
  7. 07What changed since then?

Memory graph

Nodes and edges

The organizational graph future agents will query. Nothing is stored as a transcript — everything is stored as relationships between the things the team decided.

Nodes

PeopleSessionsAgreementsTopicsDecisionsWorkshopsCompanies

Edges

AgreedDiscussedChangedConfirmedCreatedReviewedReplaced
Why am I seeing this?

Revisit fundraising philosophy

  • Three participants expressed uncertainty on capital strategy.
  • Two conflicting assumptions about dilution tolerance.
  • One unresolved agreement from the Q2 workshop.

Recommendation Agent · reading from Memory + Insight

Alignment evolution

What Memory notices over time

  • Leadership clarity improved across the last two sessions.
  • Communication expectations remained stable.
  • Fundraising philosophy shifted after Series A.
  • Decision ownership became more decentralized.

Privacy model

Designed around trust

Every agent inherits the same operating boundaries. These are product-level guarantees — not prompt suggestions.

  • Never

    Reveal private reflections

  • Never

    Quote one founder to another

  • Never

    Take sides

  • Never

    Infer personality

  • Never

    Diagnose people

  • Never

    Rank participants

  • Always

    Always synthesize across the team

  • Always

    Always explain why a recommendation appeared

  • Always

    Always preserve authorship of agreements

Developer ready

Each agent can become any of these

The architecture is intentionally decoupled. Today the agents share a runtime; tomorrow any one of them can graduate into an independent surface without changing how the rest of Coform feels.

Independent APIPrompt chainMCP toolWorkflowBackground jobVector retrieval pipeline

"We made better decisions together."The line Coform is built to leave behind