Solid AI Agents: The $6,531 Autonomous Bill
Solid AI Agents Review: Who Pays When It Loops?
Key Takeaways: What Should Developers Know Before Funding a Solid Agent?
- The number: $6,531.30. One unsupervised agent with unrestricted AWS access ran up that bill in 24 hours (May 2026, not a Solid incident).
- The product: Solid agents get their own computers, Google and Apple accounts, and budgets. Plans cost $40, $160, or $640 monthly, drawn from one balance.
- The gap: Solid's homepage and pricing page do not confirm whether that balance is a hard stop.
- The move: Start with the $20 trial, enforce approvals outside the prompt, and test the kill switch first.
What Is Solid and How Do Its Agents Get Their Own Computers, Accounts, and Budgets?
Solid gives an agent a machine, an identity, and a wallet. The founder's argument is simple. An agent that lacks a server, an account, or a way to pay stops and waits for a human.
The company says its agents run on their own Windows and macOS computers, Linux servers, iPhones, and Android phones. They hold their own Google and Apple accounts. They sign up for services and pay for them inside the budget and approval rules you set.
You describe a job in plain language. The agent plans the steps, picks the tools, and keeps working after you close your laptop. Solid says no prebuilt connector is required. An agent can call an API, build a missing integration, or drive a website, desktop app, or phone app the way a person would.
Failure handling is part of the pitch. When a tool breaks, the agent investigates, repairs its setup, and checks that the job runs again. If it stays blocked, it explains what it needs. It can also create more Solid agents or bring in Codex and Claude Code.
This is the second Solid launch on Product Hunt. The first, in 2025, ranked third for its day, when Solid was pitched as an AI builder of real web apps. At the time of writing, Product Hunt's daily list has Solid at number one.
Who Is Behind Solid and Is the $6M Funding Verified?
The founder is Trevor Keith. On Product Hunt he says the team comes from MIT, Berkeley, and Waterloo, and that Solid has raised $6M. That is a founder statement, not an audited fact.
His public profiles support the MIT and Waterloo part. An MIT and Harvard founder accelerator also lists him as Solid's CEO. Public reporting on the round is thin. No press coverage or named lead investor turned up in our search. Treat the $6M as company-reported.
One naming trap matters for anyone citing this product. Solid Data Inc., an enterprise data startup with different founders, announced a $20M seed round in February 2026. It is a different company.
The homepage also shows logos for Revolut, ElevenLabs, NVIDIA, and others. The wording says builders and operators at those companies use Solid. That is narrower than saying the companies are customers.
What Does Solid Save Developers on Setup and Environment Configuration?
The upside is unglamorous. It is setup toil.
Count the work before a job produces value. Provision a box. Create accounts. Wire OAuth scopes. Install the toolchain. Chase the version mismatch. Solid says its agents handle setup, connections, and code, even for software you have never used.
The company's example workflows show where that lands. In them, a bug-resolution agent reproduces a production alert, writes and tests a fix, and opens a pull request for engineer approval. A sales-demo agent turns meeting notes into a hosted, tested demo. An evals agent reruns after every release.
The founder adds that beta users built and deployed SaaS platforms, automated supplier calls for raw material procurement, and ran a LinkedIn recruiting pipeline.
The time savings are plausible but unmeasured. Neither the homepage nor the Product Hunt page publishes hours saved per job. Solid's own advice is sound: start with one job whose result you can verify.
What Happens When an Autonomous Agent With a Budget Hits an Error Loop?
A person who hits an error stops and asks. An agent retries.
In May 2026, an operator gave an autonomous agent full AWS access and a deadline to scan a hobbyist network called DN42. Network engineer Lan Tian documented what followed. The agent decided it needed five m8g.12xlarge instances, plus load balancers and Lambda functions. It reapplied its CloudFormation template repeatedly and duplicated the stack. The operator noticed a day later, when the card was charged $6,531.30. AWS later negotiated the bill down to $1,894.
That was not Solid. It is the failure class any autonomous AI agent with a credit card or a cloud key can hit.
Why Do Retries Turn Small Bugs Into Large Bills?
Each lap of a failing loop costs something. Model tokens. Machine time. A purchase. CloudZero says one request can trigger three to ten model calls. Multiply that by every retry. For AI coding agents, API token cost scales with laps, not prompts.
Run illustrative math. Assume 40 laps an hour at $0.75 per lap for model calls plus machine time. That is $30 an hour, or $720 a day. A $640 balance would drain in about 21 hours. These inputs are assumptions. Replace them with your own measurements.
Machines multiply the problem. Solid's pricing page says you can use as many computers and phones as the work needs, and each one draws from the balance. Solid agents can also create more Solid agents. If a bad plan gets delegated, every child can inherit it. Ask whether children share the parent's balance or carry their own.
Does One Shared Balance Cap the Damage?
Partly. Solid loads each plan payment into a balance that pays for AI, machines, and purchases. If that balance is a hard stop, the worst case is bounded by what you loaded: $40, $160, or $640 a month per plan. That beats the DN42 setup. InfoQ traced that incident to one credential that could provision anything and had no spending scope.
Three questions stay open. Neither the homepage nor the pricing page says whether the balance is a hard stop, whether it reloads automatically, or how recurring charges from agent-bought services are metered.
The earlier product offers a clue. A 2025 reviewer of the original Solid web-app builder hit a credit wall in the middle of AI error fixes. The reviewer argued that bugs caused by the system should not burn credits. That suggests a stop exists in some form. It also shows that fixing errors costs money.
Timing matters too. InfoQ reports that practitioners see cloud billing run about a day behind agent-speed spend. In both incidents it covers, the credit card flagged the problem, not AWS. Alerts arrive after the money leaves.
Which Controls Belong Outside the Prompt?
A spending limit stated in a prompt is advisory. The model can talk itself past it. Enforcement belongs in application code or an API gateway, which can refuse the next call no matter what the agent concludes.
Some of Solid's example prompts put the guardrail in plain language, such as telling the agent to ask before buying services. That works in a demo. In production, the platform must enforce the gate. Solid's FAQ says you set the budgets and approval rules, which suggests platform-side enforcement. Verify it.
Orite, a startup building spending limits for agents, makes a sharper point on Product Hunt. A budget tracks how much was spent. It does not judge whether the purchase made sense.
Here is the control stack I would want before any agent touches a card:
- A hard cap enforced outside the model. Total spend per job and per day.
- A velocity breaker. Trip on spend per minute, not just the total, so a fast loop stops early.
- Iteration and time limits. Bound the laps per task.
- Idempotent writes. Read state before writing, and key every create call so a retry cannot duplicate a resource.
- Allow-lists. Approved merchants, regions, and machine sizes.
- A tested kill switch. One action that halts every child agent and machine.
How Does Solid Compare With Chat Assistants on Cost Risk, Autonomy, and Error Recovery?
The table sets standard chat use of ChatGPT and Claude against Solid as the company describes it.
| Criterion | Chat assistants (ChatGPT, Claude chat) | Solid autonomous agents |
|---|---|---|
| Cost risk | Low. One reply per prompt. Spend tracks your plan or API calls. | Higher. Unattended jobs draw AI, machine, and purchase costs from one balance. |
| Execution autonomy | Low. You run each step and paste results back. | High. Agents pick tools, set up devices and accounts, and work after you log off. |
| Error recovery | Manual. You spot the failure and re-prompt. | Automated first. Agents investigate, repair, and rerun, then ask for help. |
| Human checkpoints | Every turn. | Only where you configure approvals. |
| Spend visibility | Plan limits or a monthly invoice. | Agent can itemize a job's AI, machine, and purchase costs. |
| Main failure mode | A wrong answer you catch on read. | A wrong action that repeats before you read anything. |
The Solid column follows the company's own descriptions, not measured behavior. The company says its agents choose tools, set up devices and accounts, repair failures, and keep working after you close your laptop. It also says the agent can itemize what a job cost across AI, machines, and purchases. The tradeoff is clean. Chat keeps you in the loop, so errors cost time. Autonomy takes you out of the loop, so errors cost money.
How Would a Senior Engineer Stress-Test Solid Before Trusting It With Real Money?
I judge agents by how they fail, not by how they demo. Anyone who has run a retry queue in production knows the pattern. The second failure is cheap. The four hundredth is not.
Here is the protocol I would run on the Starter plan.
- Fund it small. Use the $20 trial, then Starter at $40. Attach no external payment method until the controls pass.
- Break something on purpose. Revoke an API key mid-job. Return 500s from a webhook. Block a domain. Log how many laps each failure produces and whether the agent escalates.
- Reconcile the money. Ask the agent to itemize a job's cost, which Solid says it can do. Compare that number to the balance change. A mismatch is your first bug.
- Drain the balance. Let a throwaway job run dry. Check that work halts cleanly, what state it leaves, and whether machines keep running.
- Audit the leftovers. Look for orphaned machines, accounts, and subscriptions after a killed job.
- Test the approval gate. Request a purchase above your threshold. Then tell the agent to proceed anyway. The platform should still block it.
- Read the log like a postmortem. Every action needs a timestamp and a cost.
This review draws on Solid's public pages, its Product Hunt launch, and reported agent-cost incidents. It contains no hands-on benchmark numbers. Treat the protocol above as the next step, and trust your own numbers over any review, including this one.
What Are the Most Common Questions About Solid AI Agents?
Q:How much does Solid cost per month?
Solid sells three plans: Starter at $40, Pro at $160, and Max at $640. Each payment lands in full as account balance, with no platform fee. A $20 free trial is available. AI, machines, and agent purchases all draw from that balance.
Q:Can a Solid agent overspend its budget?
The design points to a ceiling. Spending draws from one balance and follows budgets and approval rules you set. The homepage and pricing page do not say whether the balance is a hard stop or can reload automatically. Test it on a throwaway job first.
Q:How is Solid different from ChatGPT or Claude chat?
Chat assistants answer inside a conversation and wait for you. Solid runs jobs on its own machines and accounts, repairs failures, and keeps working after you close your laptop. The cost of an error shifts from your time to your balance.
Q:Who founded Solid and how much has it raised?
Trevor Keith founded Solid. He says on Product Hunt that the team comes from MIT, Berkeley, and Waterloo and has raised $6M. Independent coverage of the round was not found. Solid is a different company from the enterprise data startup of the same name that announced a $20M seed round in February 2026.
Legal and financial disclaimer: This article is independent editorial analysis for informational purposes only, not legal, financial, investment, or security advice, and it relies on public information that may change after publication. To the fullest extent permitted by law, The Dollar Craft and its authors accept no liability for charges, losses, or obligations arising from autonomous software or API usage, and you bear sole responsibility for any spending authority you grant, so verify current terms and controls with the vendor and qualified professionals first.
Comments
Post a Comment