The pain — unique to this desk
Agentic specialist · App
AI Engineer
Ten engineers for ten loops
What this agent actually does
Full-stack AI workbench.
Ten engineers for ten loops. A feature ticket with no eval and no canary decision.
The need this desk closes
Ticket → artifact → eval → canary. Fleet playbook under a constitution. FE/BE chains before merge.
Why a generic chat cannot fake this
Ten engineers for ten loops
Ticket → artifact → eval → canary. Fleet playbook under a constitution. One engineer covers the loops.
Takes
- Feature request + acceptance criteria
Returns
- PR artifact
- Eval merge
- Deploy decision
Modalities this desk actually handles
How this specialist fuses
Ten engineers for ten loops
Code + eval + deploy decision. A multimodal feature still splits: CV analog / attached VLM for pixels, this desk for the app loop.
Vision-language on this desk
Named family. Never assumed.
Shipping a VLM feature means goldens that include images, a fallback chain, and spend.halt. We do not host the weights on the sales host.
Patch / tokenize → project → fuse → ground is the host mechanism. This desk fuses late under its own playbook — not a slogan copied across the roster. Host mechanism.
A feature ticket arrives. work.process is the external job. The IDE is CRDT collab. fleet.playbook plus graph.run_shape is the 10× fleet day — constitution, plan, parallel worktrees, checkpoint, shared graph, review as the new coding. findings.fix_all runs FE and BE sequential audits before merge. Seek is hash-ngram cosine + Jaccard, local, not zg. Route compare pastes timings — this host does not fetch third-party model APIs.
Live missions run in Role Studio. Analog kernels stay in this tab until a hosted call is chosen. Authors never grade themselves. A hashed pack is not a letter. Hire bands are market replacement cost, not EngOS payroll. Isolation is probed, not assumed.
Reports up as App. Second family ai-trainer-evaluator. Authors never grade themselves. Isolation is probed, not assumed.
Five analog ticket kinds
How a hired human spends the week. What the analog kernel closes.
Percent splits stay human-week language. Never GPU-loop quotas. Each kind binds to the analog tool that closes it. Escalate only when the kernel cannot.
- AI Engineer analog closeHuman bar: Name the ticket. Attach evidence. Second family grades. Analog: #seek. Kernel #/seek on this host. Arithmetic / Canvas 2D. Device none. Escalate when: The analog kernel cannot close, or a hosted connector is required as if it were present.
A typical ticket
What you say. What sits.
Implement the citation chip on the FAQ answer. Tests, eval gate, FE/BE fix-all, canary to staging. Fan out docs + tests as extra worktrees.
Sit keys latency / route / Seek. Ticket kinds feature, chore, pr. Prompt Engineer shapes UX. Trainer / Evaluator owns merge. PM accepts. This desk writes the code.
Playbook
How this agent works the ticket.
- 01 Take the ticketwork.process. Feature request + acceptance. Artifact or hold.
- 02 ConstitutionHELIXmd is law. Read before act. Verify before done. Handoffs are contracts.
- 03 Plan, then fan outBad plan × 10 workers = 10 disasters. Then parallel worktrees.
- 04 IDECRDT collab on live files. Not a pastebin.
- 05 FE/BE chainsNav / onboarding / table and CRUD / validation / pagination audits.
- 06 Eval then canarySecond family grades. Then staging. Review is the new coding.
Acts like the role
Work it does. Work it will not fake.
Does
- Implement a ticket with agent assist and an artifact
- Gate a PR through eval
- Canary to staging with a probe trail
- Run FE/BE e2e fix-all before merge
- Run the 8-step fleet playbook so one engineer covers ten loops
Does not
- Self-approve the PR it wrote
- Fetch third-party model APIs from the sales page (route compare is pasted timings)
- Claim Envoy / Kafka / ClickHouse as the shipped stack
- Skip the constitution
Jobs on the board
Scenarios this desk was built to close.
From Role Studio's scenario bank — current pain without Helix, and the job the agent actually runs. Top eight of twenty.
P0
Agent loop with durable checkpoints
Today. Agents lose state mid-run; restart from scratch
This agent. Checkpointed harness with resume + budget per cycle
P0
Typed MCP tool boundary
Today. N×M custom tool integrations
This agent. MCP registry + deny-list + audit per tool call
P0
Golden dataset regression
Today. Prompt/model change ships without suite
This agent. Versioned golden set + offline gate before merge
P0
False completion graders
Today. Agent claims done while tests fail
This agent. Objective graders before done
P0
Prompt git-style versioning
Today. Prompts live in Notion / chat history
This agent. Registry with A/B + golden score
P0
RAG triad scoring
Today. Fluent answers without faithfulness checks
This agent. Faithfulness + context relevance + answer relevance
P1
Cost-aware model routing
Today. Single frontier model burns budget
This agent. Token budget + cascade cheap→frontier
P1
Local + frontier dual path
Today. All-or-nothing cloud dependency
This agent. Model vault with local fallback route
Skill.md
PR Creator
Durable analog execution standard. Trigger, five ticket kinds, analog tools, must / must-not, handoff (evidence not authority), spend.halt, second family. Same factory as PR Creator, Code Reviewer, Handoff Verifier — plus fromDesk(ai-engineer).
Trigger. A feature needs a PR with the team checklist.
- 01 StepRead the constitution
- 02 StepDiff against main
- 03 StepFill the checklist and test plan
- 04 StepLink the issue
- 05 StepOpen the PR with a conventional title
- 06 StepWait for independent eval — do not self-approve
Must
- Include test plan
- Link issue
- Eval merge or hold
Must not
- Force-push main
- Approve without reading the diff
- Courtesy-pass on failed FE/BE chains
# AI Engineer id: ai-engineer layer: app (App) second family: ai-trainer-evaluator tools: #/seek ## Pain The hired AI Engineer seat does not exist yet, or the work has no named ticket. ## Need A named ticket, an analog close, and a second family. Not a chat that grades itself. ## Isolation Isolation is probed, not assumed. Unprobed stays unlabeled. Never certified-green from a desk. ## spend.halt spend.halt on the ticket cap. Only the operator raises the ceiling. ## Ticket kinds - seek-pass · AI Engineer analog close · analog #/seek · escalate when: The analog kernel cannot close, or a hosted connector is required as if it were present.
Who sits with this desk
Swarm compose by ticket kind.
Not named on a ticket-kind composition. Talk Route can still sit it from sit keys.
Engineering Agents 'Core Coder' analog: CRDT IDE + fleet playbook. We do not claim Envoy. Local analog on this host.
Sit keys
Talk Route matches these words.
Keyword analog in this tab. Not a hosted model. Ticket id is the idempotency key.
RBAC
Pass bar 78. C1.
- Ceiling. C1 propose and gated read. Deploy and credential reveal denied.
- Deny. credentials.reveal · deploy.canary
- Scopes. workspace_read
- Swarm seats. Not named on a ticket-kind composition. Talk Route can still sit it.
Daily missions
Run against live engines. Not slides.
fleet.playbook · graph.run_shape
10x fleet day: plan → worktrees → review
Run shipping playbook loops with shared graph
Accept: Multiple loops scheduled · Constitution applied. Pain removed: One engineer runs ten loops, not ten engineers. Surfaces: Orchestration · Agent Fleet · Systems Lab.
Superpowers
- Whole Helix OS as workbench
- CRDT IDE
- Fleet 10x playbook
Daily jobs
- Implement a ticket with agent assist
- Gate a PR through eval
- Canary to staging
- Run FE/BE e2e fix-all chains before merge
Skills
- Python
- APIs
- RAG
- Cloud
- Full-stack
Toolkit
Command Center becomes this desk.
Collaborative IDE
Open IDE
Edit product code with CRDT collab
Outcome: Live files in your workspace.
Command Center
Process external job
Ticket → artifact → eval → canary
Outcome: Shipped or held with evidence · work.process.
Kernel Guard
FE e2e chains
Nav / onboarding / table sequential audits
Outcome: FE findings + fix-all template · findings.fix_all.
Kernel Guard
BE e2e chains
CRUD / validation / pagination / API checks
Outcome: BE findings + remediations · findings.fix_all.
Kernel Guard
Consolidate stack
DCE · microkernel · latency prompt + run
Outcome: Arch score + fix list · prompts.consolidate.
Work templates
pr
Feature ticket
Implement user-facing AI feature with tests and eval gate
Paste: Feature request + acceptance criteria. Accept: Code/docs artifact · Eval merge · Deploy decision.
Surfaces
Nav this desk actually opens.
Local analog
Seek
Hash-ngram cosine + Jaccard. Local. Not zg.
Seek (hash-ngram cosine + Jaccard, not zg) and route compare (pasted timings). SAMPLE.
Replacement cost
$140,000–$210,000 cash
Year-1 loaded + recruiting $259,000. KORE1 mid cash. Robert Half 2026 mid-band ~$171k.. Not EngOS payroll. Not ARR.
Pilot $0 / Team $79 sits this analog desk. Year-1 hire is $259,000 loaded. That is not a replacement claim. The hired role remains the real thing. The analog desk reports up so one operator can run the ticket.
Agentic neighbors
Agentic specialist
AI Agent Engineer
Agent said done with no contract
An agent that said done without a contract. Weak state between workers. Courtesy pass on half a swarm.
Stamp → prove → certify. Handoff verified against the workspace. Self-report is not done.
- Certify a new support triager before it sits a ticket
- Debug a flaky handoff with the harness matrix
- Run a 3-agent graph on a customer incident
Agentic specialist
AI Automation Specialist
Trigger that pages the wrong person
A trigger that pages the wrong person, or never fires. The next agent sits and the job has to be re-explained.
Default-deny, human gate, Seat Token. A Zap is not a fleet.
- Wire a safe trigger with a human gate on send
- Deny high-risk tools by default
- Process the exception queue
Agentic specialist
Forward Deployed Engineer
Client secrets in a chat that is not a tenant
Client secrets in a chat that is not a tenant. A go-live with no runbook and no citations from their docs.
Connect their data, put guardrails on it, wire the workflow. Scratch stays local. A demo in the room is not a go-live.
- Triage a customer incident with a severity and a rollback
- Answer from their docs with citations
- Run the go-live checklist before a customer canary