Work you can pick up without doing the research first.
Tasks are generated from the model rather than written by hand, and each one arrives as a self-contained envelope carrying everything needed to execute it.
- Board
- Columns are evidence states, not team ceremony
- Assignment
- A person, a team queue, or an agent. Same envelope either way
- Repositories
- Yours. The platform is not a source control host
Five columns, and the platform enforces the gates between them.
A card does not move because someone dragged it. It moves because the evidence supporting it changed state.
Generated from the graph, with dependencies inherited. Not yet safe to start.
GateNo open blocking gaps, acceptance tests exist, and every governing decision is resolved.
Held by a person or an agent, in a workspace, under the task's declared scope.
A pull request on your own host, enriched with rules touched and blast radius.
GateDifferential replay clean, or every difference explicitly accepted by a named approver.
Counts shown are from a worked example, not live data.
The most copied artifact in the product.
Its quality decides whether agents succeed, so it is a structured object rather than a title someone has to research. Scope, callers with real traffic, governing rules, relevant schema, runtime evidence, acceptance criteria, inherited constraints and a rollback path.
The context_slice is the part that matters. It is a budgeted graph traversal assembled to fit a context window: dependency closure, governing rules, runtime evidence and relevant schema. Not a file dump, and not a keyword search.
task_id: MIG-142 goal: "Route state surcharge through TaxEngine, preserving pre-2014 New York behaviour" scope: symbols: [PremiumCalculator#applyStateSurcharge, SurchargeTable#lookup] files: [src/main/java/.../PremiumCalculator.java] tables: [RATE_TABLE_2013, ENDORSEMENT] context_slice: # budgeted graph traversal callers: [QuoteServlet 3880/d, RenewalBatch 310/d, AgentPortalAPI 20/d] rules: [RATE-0042, RATE-0044] runtime: "branch 218 true in 0.4% of calls" acceptance: characterization_tests: [GM-0091 .. GM-0138] max_diff: 0 p99_budget_ms: 400 constraints: - ADR-03 # inherited from a decision - "no behaviour change without a signed rule" depends_on: [MIG-138] blocked_by_gaps: [] rollback: { feature_flag: surcharge_v2 }
Markdown, ready to paste into any coding assistant your team already uses.
Served over MCP so external agents pull live context and post results back.
JSON dispatched to the platform's own execution layer, under the declared scope.
Dispatch is refused when the envelope is incomplete.
Not warned about. Refused. An agent working from a partial envelope produces plausible fiction, which is the failure mode the whole platform exists to prevent.
Nobody should have to rebuild a twenty-year-old toolchain on a laptop.
The workspace reconstructs the environment from the ingested build configuration. Old JDKs, .NET Framework, Oracle clients and COBOL compilers are available as pre-built images, and the task's envelope, rules, runtime evidence and acceptance tests are present the moment it opens.
Isolation is at the hardware level, because untrusted legacy code and autonomous agents share this environment and container boundaries alone are not enough. Sessions are ephemeral by default. Everything durable lives in Core and in your source control.
Merged from the dashboard, into your own repository.
Branch and pull request created from the task, with the envelope summarised in the description and linked back. Review happens inside the platform because a normal diff is missing most of what matters.
| Shown in review | Why a plain diff is not enough |
|---|---|
| Rules this change touches | A one-line edit can rewrite a business rule nobody named in the pull request. |
| Acceptance tests covering it | Reviewers otherwise have no way to tell whether the safety net exists. |
| What the replay produced | Differences by cause, with the ruling and the approver attached to each. |
| Blast radius per modified symbol | Weighted by real traffic, so the review effort lands where the risk is. |
| Reviewer assignment | Derived from ownership signals in Core rather than from a rotation. |
The audit trail runs unbroken: task, envelope, agent run, pull request, review, replay result, merge, approver. It works against GitHub, GitLab, Bitbucket and Azure DevOps, and the platform never hosts a repository of its own.