Consortio
A project-first environment for combining multiple coding-model perspectives on one repository task, with isolated workspaces, visible evidence, and human-controlled merge.
Problem
AI coding agents are useful, but serious repository work quickly becomes coordination work: preparing context, running different tools, comparing diffs, checking verification, tracking failures, and deciding what is safe to merge. A chat transcript is not enough evidence for that workflow.
Approach
Consortio treats delegated coding work as an auditable engineering workflow. A task starts as durable intent and prepared repository context, then multiple provider-backed agents can plan, build, review, judge, and verify candidate solutions in isolated workspaces. The system preserves the important evidence: intent specs, compiled intent contracts, execution briefs, plans, critiques, diffs, checks, scores, traces, proof results, verification ledgers, run reports, and export/apply artifacts.
The product direction is an artifact-backed multi-provider orchestration environment where users define intent once, watch candidate work progress, inspect the evidence trail, compare the perspectives of different models on the same task, and keep explicit control over export and merge decisions. The name reflects that benefit: a consortium of model outputs gives the operator a broader view than a single model run. Native harnesses support Codex, Claude, and Gemini directly, while OpenRouter routing allows arbitrary supported models to be brought into the same workflow.
One of the hardest engineering problems is handling conflicting model perspectives without flattening them into false consensus. Consortio preserves conflicting alternatives, marks consensus findings, tracks disagreement fingerprints, can force a full-plan rewrite when disagreement repeats, excludes candidates that remain unresolved, and records no-winner or merge-conflict outcomes as explicit artifacts.
I have designed and implemented the system end-to-end with AI agents as implementation and review collaborators. The engineering work spans headless provider drivers, PTY fallback, capability registries, quota handling, auth bootstrap, Docker sandboxing, typed schemas, artifact contracts, run graphs, context overflow handling, scoring heuristics, retry handling, test sharding, and browser-level verification of the product surface.
Outcome
Consortio's private-beta loop can run a user-defined repository task end-to-end through its command-line engine:
- accepting a stage-one intent file or run configuration
- compiling an intent contract, execution brief, and request manifest
- preparing isolated candidate workspaces and provider homes
- running builder, reviewer, judge, and verifier phases
- publishing run state, current candidate status, artifact manifest, verification ledger, and terminal run report
- exporting, applying, or merging the selected result under operator control
The repo carries docs and tests around provider routing, sandbox/auth setup, artifacts, proof planning, verification, resume/retry state, auto-merge, and nested conflict resolution.