Artifacts over chat memory
Research, positioning, and critique have explicit fields, sources, versions, and downstream consumers.
Fieldcraft architectureImplemented Phase 0
Multiple agents, one accountable operating loop.
Specialists work from structured tasks and artifacts, a Critic tests the output, and higher-risk action stops for human approval.
Implemented orchestration flow
Human missionOutcome and constraints
OrchestratorTasks and permissions
ResearcherRead-only
Research ReportTyped artifact
WriterLocal mutation
Positioning BriefTyped artifact
CriticNamed rubric
Human approvalFinal gate
If the brief does not meet the evaluation threshold, the Writer and Critic repeat before final review.
At a glance.
The operating problem.
A single general-purpose assistant can hide dependencies, sources, revisions, and authority inside one long conversation.
The user needs to know which specialist owns a task, which artifact it consumes, why it is waiting, what quality gate applies, and what the system is allowed to do next.
Before
One assistant and hidden dependencies
The brief, research, draft, critique, and decisions blur together in chat history.
After
Tasks, artifacts, evaluation, and gates
Ownership, state, revisions, permission, cost, and approval remain visible in one project record.
How the implemented loop works.
The source connects three specialists through explicit task factories and typed outputs.
- Define a missionThe Orchestrator creates and persists canonical project state.
- Decompose the workResearch, writing, and critique tasks receive an owner, expected output, and permission level.
- Build a Research ReportThe Researcher returns findings, sources, open questions, confidence, model, tokens, and cost.
- Draft a Positioning BriefThe Writer consumes the research artifact and produces a structured commercial brief.
- Evaluate with a CriticNamed thresholds test evidence, usefulness, clarity, voice, and risk awareness.
- Revise when requiredThe Writer and Critic repeat within the configured iteration limit.
- Stop for final approvalThe mission remains awaiting approval until the human accepts the final output.
The interface should make control visible.
This reference interface translates the implemented control plane into a founder-readable product model. It is a schematic, not a live product screenshot.
Shared context
OutcomeOne approved positioning brief
AudienceFounder-led service businesses
ConstraintsSource-backed claims only
Brand rulesDirect, skeptical, builder-focused
Decisions madeExternal action remains blocked
Plan and artifact workspace
Research the problemResearcher, complete, read-only
Write positioning briefWriter, complete, local mutation
Evaluate evidence and clarityCritic, review, read-only
Approve final outputHuman, waiting, external action blocked
Positioning Brief, version 2
Created by Writer from Research Report version 1. Critic score and unsupported assumptions remain attached before downstream use.
Compare versionsRequest revision
Important decisions.
Permission is part of the task
Each task declares its level, and an agent cannot execute work above its own ceiling.
Quality has a named threshold
The Critic tests five criteria instead of offering an unstructured opinion.
Final means human-approved
The loop can produce work, but the project does not become complete until the final gate is resolved.
Evidence from the implementation.
The public diagrams are faithful explanations of inspected source, not invented product screens.
Control plane
- Mission creation and task decomposition
- Researcher, Writer, and Critic dispatch
- Revision loop with an iteration limit
- Final human approval callback
Canonical state
- Mission and current phase
- Artifacts, versions, and decisions
- Approval requirements
- Token and cost accounting
What Fieldcraft delivered.
Operating architecture
Mission decomposition, agent ownership, permission levels, artifact contracts, evaluation, revision, approval, and canonical state.
Reference interface
A responsive product model for mission context, plan state, artifacts, activity, agent inspection, and approval.
Proof boundary.
Verified
- The Phase 0 Researcher, Writer, Critic loop is implemented.
- Typed artifacts, canonical state, permissions, evaluation, revision, cost and token tracking, and human gates exist in source.
Not claimed
- No production browser or GitHub integrations.
- No persistent long-term memory or autonomous external execution.
- The reference interface is not presented as a live dashboard.
More than one specialist
Map a workflow that needs more than one agent.
Define the mission, artifacts, permissions, and human gates before choosing the agents.
Map My Operating Bottleneck