Claude Agent SDK
A code agent that checks its own work before it writes.
The Claude Agent SDK is for the step the org cannot express: read several systems, reason across them, reflect on the structured output, then write once through the gateway.
Founded 2022 · Founder in the Salesforce MVP Hall of Fame · Dubai HQ · Kochi studio
What we put in
Multi-turn tool use, then a reflection gate.
The pattern we implement is simple and it is the reason these agents survive review: never write from the first draft.
01
Plan
The agent decides which tools it needs. The tool list is an allow-list, not everything the gateway can reach.
02
Call tools
MCP tools and enterprise APIs through the control plane, with the user’s identity carried forward.
03
Draft, then reflect
The agent evaluates its own structured output against a schema and the business rules before anything is committed.
04
Write once
A single idempotent write on the durable plane, with a rollback path named in the specification.
Mermaid source
flowchart TD
A["Task from a surface or a schedule"] --> B["Plan: which tools, which order"]
B --> C["Tool calls through the gateway"]
C --> D["Draft structured output"]
D --> E["Reflect: schema, totals, policy"]
E -->|"fails"| B
E -->|"passes"| F["Write once, idempotently"]
F --> G["Trace and hand-back to the desk"]When it earns the seat
Long context, messy inputs, real judgement.
Use cases
The exceptions a code agent takes.
Typical first jobs.
Example
Claims desk lead at a Dubai insurer
Today: Claims with missing paperwork bounce between three teams.
After: The agent reads the claim, policy and email trail, lists what is missing, and drafts one request. A handler sends it.
Example
Procurement lead at a Saudi manufacturer
Today: Discontinued-part requests all wait for one engineer.
After: The agent drafts a substitute from the catalogue and the spec. The engineer approves before the order changes.
Example
Contracts team at a US services firm
Today: Renewals mean reading each contract in full.
After: The agent reads the whole contract and lists every term that differs from the standard. A lawyer signs.
See how it works
Governed
A code agent is not an exemption.
The same four controls apply as to anything else on the runtime plane.
Before the write
Write from the first draft, or check it first.
A demo agent
As we build it
- Writes its first draft straight to the record
- Checks the draft against a schema and the rules
- Uses a connected app with broad access
- Calls scoped tools as the user, through the gateway
- Nobody saw its calls before go-live
- Shadow run first, write authority after
Questions
What we actually say.
- Is this the same as deploying Claude?
- No. Deploying Claude is the model decision: API, Bedrock, or Azure, with the controls your security team wants. This is the agent runtime that sits on top of whichever you choose.
- Can it write to Salesforce directly?
- It can call a scoped tool that writes, through the gateway, with the user’s identity. It does not get a connected app with broad object access.
- Which SDK would you pick for our step?
- Multi-turn tool use with reflection before a write tends to be Claude. Structured routing tends to be OpenAI. Long cyclic work with approval gates tends to be LangGraph. We classify the step first.
Next
What to read next.
The brief · one email
Send the exception the org keeps escalating.
We will tell you whether a code agent belongs on it, and what it may write. The assessment is free: a short read of the org and a written note. If you go ahead, the first build week is a fixed AED/SAR 14,900 (about USD 4,050), credited in full when you sign the build within 60 days. The honest no: if the work is a software factory with a named delivery date, we are the wrong partner.