Authoring with agents

Tarski’s authoring bet is simple: you will rarely write Tarski by hand. Coding agents — Claude Code, Codex, Devin, or any capable tool — author rules and mappers against the scenarios and invariants you declared, iterating until verification proves the contract holds. You review the contract and its evidence, never generated internals.

The authoring loop

you     declare scenarios + invariants
agent   writes .ta rules and mappers · runs tarski verify · iterates
engine  checks the corpus · every invariant · every fixpoint
you     approve the evidence

What makes the loop work is that every feedback surface an agent touches is deterministic and machine-readable:

• tarski verify is the oracle — fixture results, invariant checks, and load-time gates, ending in a machine-readable verdict on every termination path.
• Diagnostics teach. Every .ta load failure carries a stable code, exact source span, the offending construct, and a correct form. An agent’s wrong guess converges instead of thrashing.
• Golden failures render the fix. A failed expected-output comparison prints the actual output in expectation-file shape.
• The graph cuts are structural. An agent cannot wire model output into trusted facts or executable intents without an explicit decision rule — the loader rejects the program. Containment does not depend on the agent’s cooperation.

Guidance that travels with the app

Every scaffolded app and template workspace ships a version-matched tarski-expert skill under .agents/skills/ and .claude/skills/ — teaching agents the observation-first authoring model, the relay namespaces, fixtures, queries, schedules, storage, projections, and lifecycle guards in product language, matched to your CLI version.

Agent-friendly surfaces run through the whole platform:

• Code-only cloud login — tarski cloud login accepts the emailed code via --code-command, so agents can authenticate without a browser.
• A read-only MCP server — tarski mcp exposes worldview queries to dev-time coding agents over stdio, refusing to start without an explicit --actor-id and --projection.
• Fixture templates — tarski fixture template query-relay <query> mints complete agent-read lifecycles for the agent to iterate against.

Architecture feedback, not just tactics

Beyond per-rule feedback, the platform answers strategic questions with machine-checkable output: monotonicity classification labels what can react coordination-free; composition agreement rejects impossible merges at load; and tarski composition architecture proposes module splits from the ontology’s own structure — boundaries computed, not negotiated. See Workviews, Collectors & modules.

The flywheel back from production

When a contained runtime agent misbehaves, provenance pins the exact observations; the counterexample promotes into a fixture; the authoring agent iterates until it passes; verification confirms the fix. Production feeds development through one artifact chain — and the artifact chain is the review surface.