Authoring with agents
Tarski ships no coding agent of its own; it provides the contract surface, diagnostics, and machine-checkable feedback that let any capable coding agent write and check your rules.
In Tarski, your domain expertise runs as rules on a deterministic engine, and you will
rarely write Tarski by hand. Coding agents (Claude Code, Codex, Devin, or any capable
tool) write the rules and mappers against the scenarios and invariants you declare, and
they iterate until verification shows the contract holds. You review the contract and its
evidence. You do not need to review the generated rules and mappers.
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
The loop works because every feedback surface an agent touches is deterministic and machine-readable:
tarski verifyis the oracle the agent checks against. It reports fixture results, invariant checks, and load-time gates, and it ends with a machine-readable verdict on every termination path.- Every
.taload failure carries a stable code, the exact source span, the offending construct, and a correct form, so an agent’s wrong guesses converge. - A failed expected-output comparison prints the actual output in expectation-file shape.
- An agent cannot wire model output into trusted facts or executable intents without an explicit decision rule. The loader rejects any program that tries. Containment does not depend on the agent’s cooperation.
Guidance shipped with the app
Every scaffolded app and template workspace ships a tarski-expert skill matched to your
CLI version, under .agents/skills/ and .claude/skills/. The skill teaches agents the
observation-first authoring model, the relay namespaces, fixtures, queries, schedules,
storage, projections, and lifecycle guards in product language.
Other platform surfaces are built for agents too:
tarski cloud loginaccepts the emailed code via--code-command, so agents can authenticate with a code and no browser.tarski mcpis a read-only MCP server that exposes worldview queries to dev-time coding agents over stdio. It refuses to start without an explicit--actor-idand--projection.tarski fixture template query-relay <query>generates fixture templates: complete agent-read lifecycles for the agent to iterate against.
Architecture feedback
The platform also 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 computed from the
ontology’s own structure. See
Workviews, Collectors & modules.
From production mistakes to fixtures
When a contained runtime agent misbehaves, provenance identifies the observations that caused it. The counterexample is promoted into a fixture, the authoring agent iterates until the fixture passes, and verification confirms the fix. Production feeds development through one chain of artifacts, and that chain is what you review.