The pain — unique to this desk
Agentic specialist · App
AI Agent Engineer
Agent said done with no contract
What this agent actually does
Agents you can certify.
An agent that said done without a contract. Weak state between workers. Courtesy pass on half a swarm.
The need this desk closes
Stamp → prove → certify. Durable handoff verified against the workspace. False-complete graders before 'done.'
Why a generic chat cannot fake this
Agent said done with no contract
Stamp → prove → certify. Handoff verified against the workspace. Self-report is not done.
Takes
- Behavior change + acceptance tests
- ReAct / tool trace
Returns
- Live certificate
- Verified handoff
- Fleet run under a constitution
Modalities this desk actually handles
TextTool tracesCode
How this specialist fuses
Agent said done with no contract
Trace + workspace evidence. Handoff is the fusion contract: files, tests, next safe action. Images on the ticket are encoded by CV before this desk treats them as facts.
Vision-language on this desk
Named family. Never assumed.
A VLM is a tool this agent may call only if attached and allow-listed. Unattached, vision tools stay denied. False-complete graders still apply to captions.
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 worker needs to exist for a task. factory.run_line stamps, proves, and certifies it. handoff.create / handoff.verify write a contract the next agent must check against the live workspace — evidence, not authority. graph.run_shape plus graph.rules enforce typed state, budget, provenance, and trace. fleet.playbook fans out under a constitution (plan first, then worktrees, checkpoint, review is the new coding). Loop debugger, handoff mesh, assembler, and outbox are the local analogs. Autonomy ceiling C3.
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 Agent Engineer analog closeHuman bar: Name the ticket. Attach evidence. Second family grades. Analog: #loop. Kernel #/loop 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.
Stand up a support triager. Certify it. Handoff to coder is a contract. Fan out three worktrees. Do not mark done if tests fail.
Sit keys agent / tool call / loop / ReAct / handoff. Ticket kinds feature, pr, incident. Trainer / Evaluator grades. Security scans the tool boundary. This desk builds the swarm; it does not grade it.
Playbook
How this agent works the ticket.
- 01 Stampfactory.run_line. Uncertified workers do not touch customer work.
- 02 HandoffContract: objective, files, tests, known failures, next safe action. Verify against workspace.
- 03 Graph8 shapes + 7 production rules. Budget enforced. Path traced. Evidence linked.
- 04 MCP boundaryTyped tools, deny-list, audit per call. N×M custom tools are the pain this removes.
- 05 Fleet8-step playbook: constitution → plan → fan-out → checkpoint → one graph → review → nightly patch.
- 06 Grade doneFalse-complete graders. Agent claimed done vs tests. Handoff Verifier skill.
Acts like the role
Work it does. Work it will not fake.
Does
- 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
- RBAC / CRUD chain audit via findings.fix_all
- Refuse a courtesy pass on half-finished swarm work (assembler analog)
Does not
- Trust an agent self-report of done
- Run without a constitution
- Paint isolation green because the desk is C3
- Invent CrewAI / LangGraph as the runtime on this host — Helix graphs are the runtime
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
Multi-agent durable handoff
Today. Weak state between agents
This agent. Handoff contract + verify against workspace
P0
Typed MCP tool boundary
Today. N×M custom tool integrations
This agent. MCP registry + deny-list + audit per tool call
P0
False completion graders
Today. Agent claims done while tests fail
This agent. Objective graders before done
P0
Golden dataset regression
Today. Prompt/model change ships without suite
This agent. Versioned golden set + offline gate before merge
P0
Direct + indirect prompt injection
Today. OWASP #1 — injection via user and retrieved docs
This agent. Red-team pack + input/tool/output layers
P1
Episodic → lessons consolidation
Today. Same failures every week
This agent. CoALA memory + procedural skill publish
P1
Cost-aware model routing
Today. Single frontier model burns budget
This agent. Token budget + cascade cheap→frontier
Skill.md
Handoff Verifier
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-agent-engineer).
Trigger. An agent claims the task is complete.
- 01 StepLoad the handoff contract
- 02 StepCheck required outcomes against the workspace
- 03 StepCheck forbidden outcomes did not happen
- 04 StepGrade pass / fail / false_complete
- 05 StepBlock resume if mismatches exist
Must
- External evidence only
- Block if forbidden outcomes
Must not
- Trust agent self-report
- Seal until every required desk checks in
# AI Agent Engineer id: ai-agent-engineer layer: app (App) second family: ai-trainer-evaluator tools: #/loop ## Pain The hired AI Agent 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 - loop-pass · AI Agent Engineer analog close · analog #/loop · 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: this desk plus AI Engineer run the fleet playbook. We do not claim Envoy or Kafka as shipped.
Sit keys
Talk Route matches these words.
loopassembleroutboxtrace
Keyword analog in this tab. Not a hosted model. Ticket id is the idempotency key.
RBAC
Pass bar 80. C3.
- Ceiling. C3 bounded plane. Swarm and certify. Isolation is probed before it is certified.
- Deny. Empty denylist on this profile. Isolation is probed before it is certified.
- Scopes. workspace_read · workspace_write · corpus_read · corpus_write · deploy · queue · eval · incident
- Swarm seats. Not named on a ticket-kind composition. Talk Route can still sit it.
Daily missions
Run against live engines. Not slides.
graph.run_shape · graph.rules
Diamond graph with budget + provenance
Run production graph rules: typed state, budget, provenance, trace
Accept: Budget enforced · Path traced · Evidence linked. Pain removed: Invisible agent bugs become path failures. Surfaces: Orchestration · Systems Lab.
handoff.create · handoff.verify
Create durable handoff
Architect → coder contract verified against workspace
Accept: Handoff verified or blocked with evidence. Pain removed: No more 'agent said done' without proof. Surfaces: Debug Harness · Agent Fleet.
Superpowers
- 8 graph shapes + 7 production rules
- Handoffs as durable contracts
- MCP multi-server flow
Daily jobs
- Certify a new support triager
- Debug a flaky handoff with harness matrix
- Run a 3-agent graph on a customer incident
- RBAC chain audit via findings.fix_all
Skills
- Agent systems
- Memory
- Tool use
- Orchestration
Toolkit
Command Center becomes this desk.
Agent Factory
Factory line
Stamp → prove → certify a worker for your task
Outcome: Live certificate · factory.run_line.
Debug Harness
Durable handoff
Contracted state between agents
Outcome: Verified handoff artifact · handoff.create.
Orchestration
Fleet playbook
Plan-then-fan-out loops
Outcome: Parallel work with constitution · fleet.playbook.
Kernel Guard
Coding fix-all
RBAC/CRUD/agent chains + zero-omission patch template
Outcome: Coding checklist + remediations · findings.fix_all.
Work templates
pr
Agent change as PR
Ship agent behavior change only if eval merges
Paste: Behavior change description + acceptance tests. Accept: Cert valid · Eval merge · Harness green.
Surfaces
Nav this desk actually opens.
Command CenterModels & KeysRole StudioAgent FactoryAgent FleetDebug HarnessOrchestrationEval GateCollaborative IDEKernel GuardCertify & OperateUpdates
Local analog
Fan-in Assembler
No seal until every required desk checks in.
Loop debugger (repeated ReAct cycles counted), handoff mesh, fan-in assembler, ticket outbox, trace sanitizer. Ticket id is the idempotency key. SAMPLE.
Replacement cost
$160,000–$240,000 cash
Year-1 loaded + recruiting $296,000. New specialty. Priced LLM-adjacent, not a SOC code.. Not EngOS payroll. Not ARR.
Pilot $0 / Team $79 sits this analog desk. Year-1 hire is $296,000 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 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.
EventsText
- 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.
TextCustomer docsImages as attachments
- 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
Agentic specialist
AI Engineer
Ten engineers for ten loops
Ten engineers for ten loops. A feature ticket with no eval and no canary decision.
Ticket → artifact → eval → canary. Fleet playbook under a constitution. One engineer covers the loops.
CodeTextAPIs
- Implement a ticket with agent assist and an artifact
- Gate a PR through eval
- Canary to staging with a probe trail