setupaiagents.com
Free resource · tested on 40+ agent builds

The Workspace Agent Spec Template

A one-page written spec is the single highest-leverage document in any OpenAI Workspace Agent build. It forces the hard decisions before code — which is why agents shipped with one land cleanly and agents shipped without one tend to degrade in week 2. This template captures the 12 fields every agent build should answer before anyone writes a prompt or wires a connector.

free download · no spam

Get the full spec template in your inbox

Same template tested against 40+ shipped Workspace Agent builds. Use it once, reuse it forever. Plus a weekly 2-minute dispatch on what’s shipping in Workspace Agents — unsubscribe any time.

No credit card. No pushy follow-ups. One person running this — you're emailing me directly.

What’s inside

The template is structured around the 12 decisions every production agent must have answered before build begins. Skip any of them and you’ll re-do the work in week 2.

01 · Agent name and one-line purpose

What does this agent do, stated in one sentence a new team member could repeat? If you can't write it in one line, the scope is probably wrong.

Agent nameOne-line purpose

02 · Trigger — when does the agent run?

Scheduled? Event-based? Ad-hoc? Specify the exact condition. 'New lead' is ambiguous; 'new HubSpot contact with lifecyclestage = marketingqualifiedlead' is not.

Trigger typeExact conditionTimezone

03 · Inputs — what data does the agent read?

Every data source needs tool, scope, fields, and access level. Connector approvals need to be in place before the build starts.

Data sourceScopeFields/recordsAccess level

04 · Decisions / reasoning

What judgment is the agent making? 'Read the lead, write an email' isn't enough — list every decision embedded in the workflow.

05 · Outputs — what does the agent produce?

Primary output. Where it lands. Format and structure. Who sees it first. The more precise, the less rework on Day 5.

Primary outputDestinationFormatFirst reviewer

06 · Approval gates

Every external action needs a human-in-the-loop for 4–6 weeks of real use. Specify who, what, when.

07 · Success criteria

The measurable outcome that decides whether you renew, tune, or retire the agent after 30 days. Quantitative + qualitative.

08 · Cost caps

Per-agent credit cap with soft alert and hard stop. Monthly budget. Named person who gets alerted.

Soft alertHard stopMonthly budget

09 · Owner and runbook

Named internal owner (not 'the team'). Backup owner. Runbook location. Monthly health-check cadence.

10 · Kill criteria

Explicit conditions under which the agent gets retired. Agents without kill criteria decay silently and become liability.

11 · Dependencies and gotchas

Anything external the agent depends on that could break it. Known edge cases. Fallback logic.

12 · Timeline and sign-off

Spec date. Build start and completion targets. Live date. 30-day review date. Three signatures: owner, admin, builder.

How teams actually use this

  • Fill it out before every agent build — catches scope creep before code. Reviewing it out loud surfaces hidden assumptions.
  • Share with your admin for connector-approval sign-off. Section 3 (Inputs) is the bit they care about most.
  • Use it as the Day-7 acceptance criteria. If the agent matches the spec, the engagement closes cleanly. If it doesn't, the spec is the objective reference.
  • Attach to the runbook so the 30-day review has a baseline. You'll want to compare 'what we planned' to 'what the agent actually does' at the one-month mark.
  • Re-spec agents every quarter. Agents drift. The original spec plus real-world deltas tells you what needs tuning vs what needs retiring.

Want the spec filled in for you — and then shipped?

I turn the filled spec into a shipped agent in 3–5 business days for $1,000 flat. Book a 20-min scoping call and you’ll have a filled spec in your inbox within 24 hours.

Book the scoping call →See pricing

Related