AI agents
The agent permission problem
Agent projects rarely stall on model quality. They stall on the question of what the agent is allowed to do, and who answers for it.
8 min read
Practice 05
AI Agent Adoption
Put agents into real workflows without losing control of what they can reach, spend, or break.
Engagement shapes
Agent readiness review
2 weeks
Design and evaluation harness
4 to 6 weeks
Pilot to production
10 to 14 weeks
Security review of an agent estate
2 to 3 weeks
Who this is for
- Engineering leaders whose agent prototypes are blocked on security review
- Operations leaders with a workflow worth automating and no safe way to do it
- Teams running agents in production without an evaluation suite
The problem
A demo agent needs a prompt. A production agent needs an identity, a permission boundary, an evaluation suite, a cost ceiling, an audit trail, a human approval path, and a rollback.
Most projects stall in exactly that gap. The prototype convinced everyone, and then security, finance, and operations each asked a question nobody had designed for.
The failure modes are specific and known: prompt injection through untrusted content, over-permissioned tools, silent regression after a model change, and cost that scales with retries rather than with value.
What we do
The workstreams in an engagement
01
Use-case qualification
Which workflows suit agents and which are better served by ordinary automation.
02
Workflow decomposition
Broken into tools, state, and decision points before any framework is chosen.
03
Tool and API surface design
Least-privilege scoping per tool, with the blast radius of each documented.
04
Non-human identity
A permission model for agent actors, including credential issuance, rotation, and revocation.
05
Human-in-the-loop design
Approval gates, escalation, and reversibility, placed where the cost of a wrong action is highest.
06
Evaluation harness
Task suites, regression sets, and scoring you can trust, handed over so you can re-run it on every change.
07
Observability and tracing
Runs, spans, tool calls, and failures, linked to cost and latency per task.
08
Cost controls
Budgets, rate limits, model routing, and caching, set before the first production run.
09
Security review
Prompt injection, data exfiltration, over-permissioned tools, untrusted content handling, and supply-chain risk in agent frameworks.
10
Pilot build and hardening
One workflow taken to production with runbooks and an on-call plan.
Deliverables
What lands on your desk
You keep all of it, including the method behind it, so the work can be repeated without us.
- Agent opportunity map with qualification rationale per workflow
- Reference architecture and tool-surface specification
- Evaluation suite and baseline scores, handed over and re-runnable
- Guardrail and permission specification
- Observability and cost-control design
- A working pilot with runbooks and an on-call plan
Standards we work against
- OWASP Top 10 for LLM Applications
- NIST AI Risk Management Framework
- NIST SP 800-207 zero trust principles
What we do not do
- We do not ship an agent with production write access and no evaluation suite, regardless of deadline.
- We do not build an agent for a workflow that a scheduled job would do better.
Engagement shapes
Three sizes, not one package
Durations are indicative and depend on estate size. Scope and price are fixed in writing before the engagement starts.
| Shape | Duration | What you get |
|---|---|---|
| Agent readiness review | 2 weeks | Qualification, blockers, and a design direction |
| Design and evaluation harness | 4 to 6 weeks | Architecture, guardrails, and a working eval suite |
| Pilot to production | 10 to 14 weeks | One workflow live, observable, and owned |
| Security review of an agent estate | 2 to 3 weeks | Findings, permission remediation, and an eval gap list |
Related insights
Our thinking on this
AI strategy
Why AI pilots stall between demo and production
The gap is rarely model capability. It is unit cost, data readiness, evaluation, and an unanswered question about accountability.
7 min read
Next step
Start with a briefing, not a proposal.
Thirty minutes on ai agent adoption. We will tell you whether we are the right firm for the problem, and who to talk to if we are not.