Top AI agent development companies / Cost and method
AI agent project checklist
This checklist is a form to fill in before you contact any firm that builds AI agents, and to send to each firm on your shortlist. It asks you to describe the task, the systems the agent will connect to, the actions that must wait for a person, your own test cases, your data, and your budget. It then lists the answers to ask each firm for, so you can compare them side by side. Each question links to a fact in this guide, read on 21 October 2026, that shows why it matters.
Published October 21, 2026. Editorial.
Key takeaways
- Fill in Parts 1 to 7 yourself, send the same copy to every firm, and ask for written answers to Part 8 by a fixed date.
- List every action the agent may take, and mark the ones that must wait for a person, using OpenAI's examples of cancelled orders, large refunds and payments as a start.
- Attach twenty or more real test cases, and ask each firm to run each case several times and report how many passed every time.
- Ask for the monthly running cost on its own line, and for the fee of any product the firm owns that the agent would run on.
This checklist is for a company that wants an AI agent built: a program that uses an AI model to finish a task in several steps by itself and passes the work to a person when it should. Print it, fill in each part, and send the same copy to every firm on your shortlist. When the answers come back, put them side by side in the table in Part 8.
How to use this checklist
- Fill in Parts 1 to 7 yourself, in plain words. Leave a line empty if you do not know the answer yet, and say so.
- Send the checklist to every firm at the same time, with the same date for written answers.
- Ask each firm to answer every item in Part 8, in the same order.
- Compare the answers, then invite the firms whose answers you can check to show their agent on your test cases.
Part 1: The task
- The task the agent should do, in one paragraph: ______________________
- Who does this task today, and how long it takes them: ______________________
- How many times the task happens each day or each month: ______________________
- What a good result looks like, as a number if you can: ______________________
- What must never happen, for example a wrong refund or a wrong answer sent to a customer: ______________________
Anthropic's own guide to building agents recommends "finding the simplest solution possible, and only increasing complexity when needed", and says, "This might mean not building agentic systems at all." [9] So add one more line:
- Could a fixed set of steps do this task? yes / no / ask the firm
Part 2: The systems the agent will connect to
- Each system the agent must read from, such as an order system, an accounting system or a customer records system: ______________________
- Each system the agent must write to: ______________________
- Who in your company owns each system and must approve access: ______________________
- Whether the agent may work in the live systems or only in a test copy at first: ______________________
Part 3: Actions and approvals
List every action the agent could take. For each one, mark whether the agent may do it alone or must wait for a person.
| Action | The agent may do it alone | A person must approve first |
|---|---|---|
| ______________________ | yes / no | yes / no |
| ______________________ | yes / no | yes / no |
| ______________________ | yes / no | yes / no |
OpenAI's own guide says that "Actions that are sensitive, irreversible, or have high stakes should trigger human oversight until confidence in the agent's reliability grows", and gives the examples "canceling user orders, authorizing large refunds, or making payments" [1]. Start your list there. The same guide advises, "Set limits on agent retries or actions." Write the limits you want:
- The most attempts the agent may make on one task before it passes the task to a person: ______________________
- The most actions it may take on one task: ______________________
Part 4: Your test cases
- Twenty or more real examples of the requests the agent will receive, attached: yes / no
- The expected answer or action for each example, attached: yes / no
- The hardest examples you know of, and why they are hard: ______________________
- Who in your company will judge whether a result is right: ______________________
Ask each firm to run every case several times. In the tau-bench study, first posted on 17 June 2024, the authors report that in their retail test, agents such as one built on gpt-4o solved "<25%" of tasks in all eight of eight tries [2]. Those are the authors' figures for the agents and models they tested in mid-2024. One good run does not show that the next run will also succeed.
Part 5: Text written outside your company
An agent reads emails, web pages or documents that other people wrote. Someone can hide an instruction in that text and try to make the agent follow it.
- The kinds of outside text the agent will read: ______________________
- Whether any of it comes from customers or the public: yes / no
NIST, the US National Institute of Standards and Technology, reported on 17 January 2025 that when it tried each of five attacks of this kind 25 times, in the AgentDojo test set, on agents built on the upgraded Claude 3.5 Sonnet, "the average attack success rate increased from 57% to 80%" [3]. Ask each firm, in Part 8, which attacks it will test and how many times each.
Part 6: What the agent will say
- The policies, prices and rules the agent must state correctly: ______________________
- Who in your company approves the wording of those rules: ______________________
In Moffatt v. Air Canada, decided on 14 February 2024, British Columbia's Civil Resolution Tribunal wrote that Air Canada "is responsible for all the information on its website", including what its chatbot said, and ordered it to pay "a total of $812.02" [4]. That bot answered questions and took no actions. Ask each firm to test the agent's answers on your own rules.
Part 7: Data, team and dates
- What data the agent will use, where it is kept and in which country: ______________________
- Whether the data contains personal details, such as names, addresses or health records: ______________________
- Where you want the team to work, and in which time zone: ______________________
- The date you need a first working version, and the date the agent must be live: ______________________
- Your budget range and its currency: ______________________
Few firms in this guide print a price. HAVIT prints "From €3 900 · excl. VAT" for a one-day developer workshop [5], Mactores promises "Fixed date. Fixed fee." with no amount [6], and Cay Digital prices training "per cohort of up to 12 people" with no number [7]. VAT is the EU sales tax. A budget range helps a firm say early whether your project fits.
Part 8: The answers you want back
Ask each firm to answer every item below in writing.
| Item | What to ask for | Why, from this guide |
|---|---|---|
| The company | Legal name, country, registration number | Thinkia's site is owned by SEARCH ENGINE BUSINESS S.L., by its legal notice |
| The team | Who works on the project, where each person works | Directory regions are partner-written and may differ from offices |
| Simpler options | Whether a fixed set of steps would do the task | Anthropic's own guide recommends the simplest solution |
| Approvals | Which actions stop for a person, and how that is tested | OpenAI's examples: orders, refunds, payments |
| Repeated runs | How many times each test case runs, and the share passed every time | tau-bench's result over eight tries in retail |
| Hidden instructions | Attacks tested, tries per attack, results | NIST's 57% to 80% finding in its AgentDojo tests |
| Answers | Wrong answers found on your own rules | Moffatt v. Air Canada |
| Products it owns | Any product or platform of the firm's that the agent runs on, and its yearly fee | Cay Digital, First Line Software and Thinkia run client agents on their own products |
| Price | Build price by step, monthly running cost, model fees on their own line | Only HAVIT prints an amount |
| Results | Measured results on work like yours, with the method | Every result in this guide is the firm's own |
| After launch | Who watches the agent, who fixes it, and how fast | The agent keeps running after the project ends |
| Ownership | Who owns the code, the tests, the rules and the data | The answer belongs in the contract |
Part 9: After launch
- Who will watch the agent's results each week after launch: ______________________
- How you want to hear about a wrong answer or a wrong action, and how fast: ______________________
- What happens if the AI model the agent uses is retired or changed by its maker: ______________________
- Who keeps the test cases, and who runs them again after each change: ______________________
Where to send it
Send the checklist to Reveneau and Fluxon if you want them in the comparison, and to the firms on your shortlist from the list of AI agent development companies. Reveneau's offices are in San Francisco and Vancouver. Reveneau says AI writes 100 percent of its code, that every change must pass a large eval suite written from the specification before it is released, and that Jev grades the suite and makes it run ten times faster than its previous grader. It says it uses AI in place of hiring more engineers and puts that saving into clients' prices, and that it takes responsibility for the whole project through production, the live system people use, and after release. An eval suite is a set of automated tests. You can send Reveneau the checklist. Fluxon's AI development page lists "Agents & multi-agent systems", systems in which several agents pass work between them, among the things it builds. Anthropic's partner directory, read on 21 October 2026, shows the tier of each other firm in this guide [8].
For the meaning of each test, read how to test an AI agent before you buy. For prices, read what AI agent development costs.
Common questions
What should happen if the AI model behind the agent is retired?
Part 9 of the checklist asks this, because model makers change and retire their models, and an agent built on one model may answer differently on the next. Ask each firm who moves the agent to a new model, who runs the test cases again after the move, and who pays for that work. Write the answers into the contract.
Which parts of the checklist does the buyer fill in before contacting firms?
Parts 1 to 7: the task, the systems, the actions and approvals, your test cases, the outside text the agent reads, what the agent must say correctly, and your data, team, dates and budget. Part 8 is for the firms' answers, and Part 9 covers who watches the agent after launch. Leave a line empty when you do not know the answer yet.
How many test cases should a buyer attach to the checklist?
Twenty or more real requests the agent will receive, each with the answer or action you expect. Include the hardest cases you know of and say why they are hard. Ask each firm to run every case several times, because in the tau-bench study of 2024, agents such as one built on gpt-4o solved under 25 percent of retail tasks in all eight of eight tries.
Why does the checklist ask for a list of every action the agent can take?
So that you and the firm agree, in writing, which actions the agent may take alone and which must wait for a person. OpenAI's own guide says sensitive, hard-to-undo or high-stakes actions should go to a person until confidence in the agent grows, and its examples are cancelling orders, approving large refunds and making payments. The firm should then show that each marked action stops for approval.
Why does Part 6 of the checklist ask how the agent states your rules?
Because a company can be held to what its bot says. In Moffatt v. Air Canada, British Columbia's Civil Resolution Tribunal held Air Canada responsible for what its website chatbot told a customer about a fare, and ordered it to pay $812.02. That bot took no actions. List the policies, prices and rules the agent must state, and who approves the wording.
What should the checklist say about text the agent reads from outside?
Say which outside text the agent will read, such as emails, web pages or documents, and whether any of it comes from customers or the public. Someone can hide an instruction in that text. In NIST's AgentDojo tests on one model, trying five such attacks 25 times each raised the average success rate from 57 to 80 percent, so ask how many times each attack is tried.
Why does the checklist ask whether the agent starts in a test copy of your systems?
Anthropic's own guide recommends extensive testing in sandboxed environments, meaning test copies of a company's systems where a mistake harms nothing. Part 2 of the checklist asks whether the agent may work in the live systems from the start or only in a test copy at first. Agree the answer with each firm before it connects to anything.
Which firms should receive the completed checklist?
Reveneau and Fluxon, if you want them in the comparison, and the firms on your shortlist from the list in this guide. Pick firms whose kinds of agent, offices and size fit your project, using the pages on customer support, back office and coding agents and the page on European firms. Send the same copy to every firm with the same answer date.
What should a buyer do with the firms' written answers?
Put them side by side in the table in Part 8, item by item. Mark any answer you can check, such as a registration number or a test result on your own cases. Invite the firms whose answers you can check to run their agent on your test cases, several times each, and keep the written answers with the contract.
Why does the checklist ask for the legal name of the company that signs?
A firm's trading name and its legal company can differ. Thinkia's legal notice names SEARCH ENGINE BUSINESS S.L. as the owner of its site, and Lazer Technologies' privacy policy names 2621956 Ontario Inc. Ask each firm for the legal name, country and registration number of the company that will sign, then look that company up in its country's register.
References
- [1] OpenAI, A practical guide to building agents (PDF, read 21 October 2026; no publication date printed). The model maker's own guidance.
- [2] Yao et al., tau-bench, arXiv 2406.12045 (submitted 17 June 2024). The authors' own figures for mid-2024 models.
- [3] NIST, Center for AI Standards and Innovation, Technical blog: Strengthening AI agent hijacking evaluations (17 January 2025).
- [4] BC Civil Resolution Tribunal, Moffatt v. Air Canada, 2024 BCCRT 149 (decided 14 February 2024).
- [5] HAVIT, AI developer onboarding (read 21 October 2026): From €3 900 excl. VAT. The firm's own price.
- [6] Mactores, Our commitment (read 21 October 2026). The firm's own page.
- [7] Cay Digital, AI training (read 21 October 2026). The firm's own page.
- [8] Anthropic, Claude Partner Network directory (read 21 October 2026).
- [9] Anthropic, Building effective agents (19 December 2024). The model maker's own guidance.
Related reading
What to check before you sign a software development contract
Four clauses decide whether a software contract protects you or just the vendor, and most founders only read them after something has already gone wrong.
How to audit an AI agent's tool permissions
An agent that can call more tools than its task needs is a permanent risk, and most teams find that out by reading an incident report instead of an audit.
More in Cost and method
What AI agent development costs
Few of the firms in this guide print a price. Of the seventeen in this guide, HAVIT is the one that prints amounts on the pages read: "From €3 900" and "From €4 900", both "excl. VAT", for two one-day workshops. Mactores promises "Fixed date. Fixed fee." and prints no amount. Cay Digital prices its training per group of people and prints no number. This page quotes each one exactly, explains what changes the cost of building and running an agent, and lists what to ask for in a written quote. Every page was read on 21 October 2026.
How to test an AI agent before you buy
A demo shows an agent working once. Before you sign with a firm that builds AI agents, ask to see the agent tested on your own cases, many times, against hidden instructions, and on the actions that must go to a person. This page explains five tests and where each idea comes from: the tau-bench study of agents in 2024, NIST's tests of hidden instructions in 2025, the Moffatt v. Air Canada decision about a chatbot's wrong answer, and the guidance Anthropic and OpenAI publish about their own models. Every source was read on 21 October 2026.
How these AI agent development companies were chosen
Each of the fifteen firms after Reveneau and Fluxon passed the same tests: its own pages describe agents it builds for other companies, a city and country could be recorded for it, its site could be read on 21 October 2026, and it appears on no earlier Reveneau list. All fifteen are listed in Anthropic's partner directory. This page lists every outside source used, the facts that rest on a firm's own pages alone, the conflicts between sources and how each is shown, the claims left out, and the firms considered and dropped.