Optional fractional seat
Agent Architect
Owns how your own agent reaches outside tools, and how those connectors sit beside the integrations you already run. Posted as an architect in some companies and a product manager in others; it’s one job.
If you hired for this full time
You would post for
- AI Architect
- Principal Engineer and Architect, Agent Platform
- Product Manager, AI Agents and MCP Tools
- Product Manager, Enterprise Agent Platform
The job usually covers
- Design the agent architecture: harness, tool use, memory, guardrails, evaluation, observability
- Own the tool surface: descriptions, gaps in coverage, what gets called and what fails
- Connect the agent to outside systems, and decide what’s native, self-serve or customer-added
- Design credentials and isolation: token brokering, least privilege, audit trails
- Keep third-party connectors working: catch upstream breaks before customers do
- Write the decision records and reference patterns other teams build on
- Evaluate before launch and monitor after
- Choose models by capability, latency and cost per task
What we take on
We own
- The map of every way your agent reaches a third party today
- The decision record for how integrations, your MCP server and external MCP servers coexist
- Tool registry design: naming, precedence and filtering
- Credential design for both token stores
- The evaluation set for tool use
- The path from "the agent did this once" to "this now runs as a workflow"
- The connector watch: upstream changes and authentication failures
We contribute, you decide
- Per provider: integration only, connector only, or both with a stated order
- Which tools are on by default
- Which outside servers are listed, and whether customers may bring their own
- Guardrail policy
- Model choice
Stays with you
- Security sign-off
- Your agent’s voice and user experience
- Model and vendor contracts
- Pricing of agent usage
Most teams are stuck at "we don’t know what to do". The work is getting that down to three or four decisions per provider with the trade-offs written out. We can build what follows because the integration side is already our trade.
What we need access to, and why
| Access | Why |
|---|---|
| Agent codebase, read; tool layer, write | The registry and tool definitions are code |
| Integration repository, write | The two sides have to be designed together |
| Agent traces and tool-call logs, read | To see what’s called, what fails and what’s never used |
| Credential architecture, not the secrets | To design the second token store without touching customer tokens |
| A staging tenant with connected accounts | Precedence and conflicts only show up with real connections |
| Your security contact | Every new outbound path is reviewed |
Agent codebase, read; tool layer, write
The registry and tool definitions are code
Integration repository, write
The two sides have to be designed together
Agent traces and tool-call logs, read
To see what’s called, what fails and what’s never used
Credential architecture, not the secrets
To design the second token store without touching customer tokens
A staging tenant with connected accounts
Precedence and conflicts only show up with real connections
Your security contact
Every new outbound path is reviewed
How you’ll know it’s working
- Tool-call success rate
- Customers with at least one tool connected
- Time to add a connector
- Conflicts between an integration and a connector reported by customers
- Agent tasks promoted to workflows
The first 30 days
- Map every way the agent reaches a third party
- Write the decision record with the options per provider
- Get one provider working end to end with both an integration and a connector under one listing
- Run a baseline evaluation
The same seat in each lane
Grow
The connector directory inside your product, including bring-your-own MCP and partner-listed connectors.
The Grow laneTransform
Your product’s own agent: how it reaches outside tools, next to the integrations that already exist.
The Transform laneHow we join
Alongside
- When
- Someone owns the function and is stretched thin.
- What we do
- We take the operational half. They keep strategy and relationships.
- How it ends
- Scope shrinks or grows by agreement each month.
Interim
- When
- The role is open or not approved yet.
- What we do
- We cover the seat day to day, reporting to whoever the hire will report to.
- How it ends
- Handoff documents and a warm transition to your hire.
Embedded
- When
- You want one accountable person without adding headcount.
- What we do
- Your email, your Slack, your onboarding. Paid through one contract.
- How it ends
- A notice period, then the same handoff as interim.
True of every seat
Behind the seat
- One named person holds the seat and is the only face you and your partners see.
- Our delivery team stands behind them, so a recommendation can become a shipped build.
- Where we bring tools of our own, or set up an integration factory, they live in your repository.
- Agents draft and the seat holder decides. Nothing reaches a partner or a customer without a person sending it.
Access rules
- Every access request is tied to a task. If we can’t name the task, we don’t ask.
- Named accounts only, on your single sign-on where you have it.
- Read access first. Write access is requested per system when a task needs it.
- Everything is revocable in one step, and the handoff checklist includes revoking it.
Rhythm
- A weekly written update: what moved, what’s stuck, what we need from whom.
- One standing meeting with the person who owns the function. Everything else is async.
- A monthly review against the measures, where scope is adjusted.
- A decision log that you own, so the reasoning outlasts us.
Guards
- Non-exclusive. You can hire for the role at any time.
- A notice period in both directions.
- Handoff documents are a deliverable.
- Your right to take the work in-house is written in, including any code in your repository.