Playbook · Building with AI

Updated

What does an AI agent developer build, and do you need an agent?

An AI agent developer builds software where a language model picks its own next step and calls tools until a task is done. Most products don’t need that. If you know the steps in advance, one model call or a fixed chain of calls is cheaper and easier to test.

Anthropic, which builds Claude, advises “finding the simplest solution possible, and only increasing complexity when needed”, and says this “might mean not building agentic systems at all.”
Source: Anthropic, Building effective agents, 19 Dec 2024

From my work

At Peeps, we looked at building a social network where gamers share what they record. The competitors all had one. We chose to let gamers share to the social networks they already use, which got them the same result with far less to build.

Agents are often the same kind of decision. Others having one is a weak reason to build one. I ask whether it is the best way to get your users the result.

One call, a workflow, or an agent

  • One call: text goes in, an answer comes out. Translating, summarizing, sorting a ticket, pulling fields out of an email.
  • A workflow: in Anthropic’s words, “LLMs and tools are orchestrated through predefined code paths.” Your code decides the steps. The model fills in some of them.
  • An agent: the model directs its own process and tool use. Worth it when the steps can’t be known up front, like a support case that needs lookups in three systems before anyone knows what the problem is.

When a plain API call is enough

Translation is a good example. The input is a message and the output is the same message in another language. EZ Chat, my own app live on the Overwolf store, translates chat between players with the OpenAI API. A message has no next step to plan, so an agent loop would only be a slower way to spend more.

  • You know the shape of the input and the output.
  • One answer per request is the whole job.
  • A wrong answer can be caught by a check in code or by a person.
  • You need the same speed and cost every time.

What an agent costs to run

An agent sends the whole conversation back to the model at every step, so the bill grows with each step. Picture a meeting where everyone rereads the full minutes before speaking. Here is my own example math, using OpenAI’s price for gpt-6.1-terra: $1.00 per 1M input tokens and $5.00 per 1M output tokens (seen 17 Nov 2026).

One call with 2,000 tokens in and 500 out costs about $0.0045. An agent run with 8 steps, where the context grows to around 20,000 tokens, sends about 96,000 tokens in and 4,000 out. That is about $0.12, or roughly 25 times the single call. Cached input, at $0.05 per 1M tokens for that model, brings it down, but not to one call.

Model choice moves this more than anything. On the same page, gpt-6-luna costs $0.10 in and $0.50 out per 1M tokens, and gpt-6-astra costs $10.00 in and $50.00 out. Anthropic says it plainly: agentic systems “often trade latency and cost for better task performance.”

What makes an agent safe to ship

  • Tools with the smallest permissions that work. Read-only first.
  • A person approves any action that spends money, sends a message or deletes data.
  • A step limit and a cost limit on every run. A stuck agent never gets bored and stops on its own.
  • A log of every step, so you can see why it did what it did. You will want to ask.
  • Tool arguments checked in code, because a model can call a real tool with made-up values.
  • A fixed set of test tasks, rerun after every prompt or model change, in a sandbox before production.

Frameworks

Anthropic warns that agent frameworks “often create extra layers of abstraction that can obscure the underlying prompts and responses.” Start with the provider’s API directly. Add a framework once you can explain what it does for you without opening its docs.

How I help

Bring me the agent idea and I will tell you if it should be an agent, a workflow or one call, sketch the design, and estimate the cost per run from the provider’s price page. The Limited plan is $1,995 a month: one task at a time, a day of work delivered within 2 to 3 working days, messages answered the same working day and calls by appointment. Advice, reviews, fixes and build work are all tasks. I sell this, so read it with that in mind.

If you only need one AI feature added to an app you already have, it is done on a monthly plan, one task at a time: Limited at $1,995 a month or Full at $4,995. Both are a flat monthly price, pause or cancel anytime. On the Full plan it usually ships within the first month.

Research into new agent designs, or a model trained for you, is a job for an ML research team.

Read next

Sources

Start

Find out what stands between you and launch.

Start the 14-day Trial: your problems ranked in the first 5 business days, then fixes for what fits in the 14 days. Not ready yet? Ask for a free teardown.

Start the 14-day Trial · $995

Secure checkout via Stripe · Refund policy

Founding rate: the first 3 clients get 20% off, in return for a named testimonial and a case study. On a monthly plan, the 20% applies to the first month.

Prefer to talk first? Book 30 min or email badii@badii.studio

Free teardown

48h · max 5 / week

A 5 to 10 min recorded review of your live app's public surface. No access needed, just the URL.

I use these details only to reply and record your teardown, and delete them after 12 months. Privacy