Corporate project · Presented company-agnostic: architecture and decisions only.
Persisted state machine coordinating AI agents from backlog item to approved pull request
A ready signal creates a persisted run; the orchestrator dispatches definition, QA planning, plan approval, development, review and rework, then publishes a pull request, watches CI, and waits for human approval.
Click a step to jump to it. Click a component for details.
An internal platform that turns work-tracking items into reviewed pull requests by coordinating specialized AI agents for requirements, QA planning, development, adversarial review and CI repair. A multi-provider LLM gateway with a prompt-safety layer, an MCP server exposing scoped tools, vector memory and a code-to-graph indexer provide the shared infrastructure, with human approval gates at every critical step.
Teams wanted LLM-assisted delivery but could not allow agents to talk directly to providers, leak sensitive data in prompts, hold unrestricted repository or backlog access, or lose work when a long-running run crashed. The challenge was to make agent behavior auditable, resumable and governed, rather than a chain of ad hoc prompts.
01
Workflow state, tools and memory live in the orchestrator and agent runtimes; the gateway only provides safe, audited, routed model access and knows nothing about work items or branches. This keeps each concern replaceable and the gateway reusable by any caller.
02
LLM runs are long, flaky and expensive. Persisting the phase, pending work and execution lease allows recovery after crashes, scheduled retries on provider failures, and parking runs that need a human decision, then resuming the same transcript and plan.
03
Scanning with Cerberus before routing means sensitive content never reaches an external provider, with no extra network hop. Rules are shared data, so they evolve without redeploying the gateway.
04
An adversarial reviewer on a different role, and preferably a different model, reduces correlated blind spots. Review findings flow back to the dev agent as structured rework, with a bounded number of iterations before escalating to a human.
05
Permissions per agent role, protected-branch rules and branch naming are validated server-side, and no generic mutation tool exists. A compromised or confused prompt cannot exceed its authority.
06
Stable symbol identities (hash of repo, qualified name and signature) and per-commit generations make context retrieval reproducible and cheap, and let multi-solution repositories merge into a single index.
Project story
Agentic Software Delivery Platform
Rotate your phone or open fullscreen to watch the story.
Open fullscreenClick or use the arrow keys to step through