Forward Deployed Engineer, how I deliver it, requirement 4 of 12
AI agents and orchestration
I build agents that call your systems through registered tools, still in early access, and workflows that start on business events.
The requirementBuild agentic AI solutions, with hands-on experience in AI agents and workflow orchestration.
How I deliver it at a customer
- Scope the jobI write down the task, the systems it touches, and each action the agent may take. I mark every action that changes a live system, since those will wait for a person to confirm them.
- Wire the toolsI register each action as a tool in Luca Gatekeeper, with a JSON Schema for its arguments and the capability a caller needs. Gatekeeper refuses a call when the caller lacks that capability or the arguments break the schema. It writes each refusal to its action log.
- Stand up the agentI start each agent as a session in the agent runtime, which runs the task through the Claude Agent SDK. The runtime saves each session to Postgres with its model, status and token total, and can reload it after a restart. Its HTTP endpoints are not reachable yet, so this part is early access.
- Set the guardrailsI flag each action that changes a live system as destructive in Gatekeeper. A destructive call waits as a pending confirmation, which expires after five minutes by default. It runs only after the human user who started the request approves it, and an agent or service account cannot approve it. Gatekeeper logs the pending request, the person's denial or its expiry, and the run after approval. This part is early access.
- Orchestrate the workflowI wire the CRM sales automation engine to start workflows on events such as lead created and deal stage changed. It runs the steps in order, skips a step whose condition fails, and stores each step's result, error and duration. A callback step is saved as pending, but nothing places those calls when they come due yet.
- Record what ranI check that the Gatekeeper action log, the agent sessions and the workflow step results are written to Postgres. Admins can query the action log, and each workflow run returns its step results, so the team can see what ran. Pod logs go to the platform's Grafana and Loki stack, and no metric for agent sessions or Gatekeeper tool calls is exported yet.
Tools I use
| Tool | What it does here | Status |
|---|---|---|
| Agent runtime and tool bridgeEarly access | Runs agent sessions through the Claude Agent SDK and checks tool permissions and timeouts; its HTTP endpoints are not usable yet. | Early access |
| Luca GatekeeperEarly access | Registers each tool with an argument schema and a required capability, holds destructive calls for a person to confirm, and keeps an action log. | Early access |
| Policy engine command rulesEarly access | Labels each shell command allow, confirm or block from a fixed pattern list, with DROP DATABASE and git push --force to main labeled block. | Early access |
| CRM sales automation engineAvailable | Starts workflows on CRM events, runs their steps in order with conditions, and stores each step's result, error and duration. | Available |
| Cloudflare Workers AI demo agentDemo only | Runs the public demo on this page with Llama 3.3 70B and three demo tools, and calls no Luca or customer system. | Demo only |
Available means shipped and in use. Early access means built but not generally available. Roadmap means designed, not built. Demo only means it powers the public demo, not customer deployments.
Ask a question and watch a tool-calling agent choose among three tools, with at most four calls. Every step shows the tool, its arguments and exactly what the model saw, then the final answer.
Open the live demoThe other answers
LLM applications in productionOn Luca, model calls normally go through a gateway that holds provider keys, records each chat call and sets a model per task.Prompt engineeringI write the output shape first, measure the prompt against each model's context window, and check replies against that shape.Retrieval-augmented generationChat answers over a customer's own documents, returned with the matched document names, after a check that the caller may use that persona.Request classification and routingI build automation that classifies each inbound request, plans bounded tasks, and holds high-risk tasks for approval; the dispatcher is in early access.AI workflow orchestrationI map the steps, API contracts, and data flows between a customer's systems and the models before I build the workflow.API and integration designI publish the API contract first, gate it with keys and scopes, then build integrations against it.Hybrid enterprise integrationI connect AI services to your identity, applications and data, and run the platform on Kubernetes on Azure Local.Prototype to productionI take a prototype through generated tests and a two-approval merge gate to a production deploy that rolls back if the new pod keeps crashing or cannot start.Approvals and exception handlingIn chat, I hold each Gatekeeper tool marked destructive until the signed-in user approves it, with expiry and a record. This path is in early access.Evals and output validationI built a versioned rubric judge for a tutoring app, and I trace failed or slow model calls to their provider.Client-facing deliveryI turn engineering work into updates that client architects, security leads and business owners can each act on.
Built by Gus Assuncao of Gus IT LLC for a Forward Deployed Engineer application. Luca is the AI platform Gus IT builds and operates. Contact gus@gusit.de. Claude is a product of Anthropic PBC. Cloudflare and Workers AI are trademarks of Cloudflare, Inc. Llama is a trademark of Meta Platforms, Inc. Other names are trademarks of their respective owners. Gus IT LLC is an independent engineering firm and is not affiliated with or endorsed by these companies.