← All articles
AI Agents3 min read

The second agent is harder than the first

The first agent is a project. The second is a platform problem — shared evals, routing, and ownership are what keep a growing agent fleet from becoming a mess.

The second agent is harder than the first

The first agent is a project. The second agent is a platform problem wearing a project costume. The first one gets the good prompt, the careful evals, the senior engineer who checks its outputs every morning. The second one gets built by a different team, on a different week, with a different idea of what "good" means — and suddenly you have two agents that disagree about the same customer, call the same tools in different ways, and fail in ways nobody can compare.

The mistake is treating each agent as its own little product. What you actually need after the first one is shared ground: one way to log what agents do, one place to see what they cost, one set of rules about which tools can touch money and which can only read. Without that, every new agent re-invents the plumbing, and every incident becomes an archaeology dig across three codebases and two chat histories. The teams that scale past three agents aren't smarter — they stopped rebuilding the floor every time they added a room.

Start with the boring shared pieces before you build agent number two. One eval harness every agent runs against, so "it works" means the same thing in every team. One routing rule for which agent handles what, written down where a human can read it — because two agents answering the same question differently is worse than one agent answering it slowly. One kill switch pattern, so freezing one agent doesn't mean learning a new dashboard in the middle of an incident. None of this is glamorous. All of it is what keeps agent number five from taking down agent number one.

Then decide who owns the fleet. Not who owns each agent — who owns the spaces between them. The shared tools, the overlapping prompts, the customer who got two different answers from two agents in the same afternoon. If nobody owns the gaps, the gaps own you. The usual shape that works: each agent has a named owner, and one person owns the platform — the logging, the evals, the tool contracts — even if that's a fifth of someone's week. A fleet without a platform owner drifts into five snowflakes that share nothing but a vendor bill.

The test is simple. Can you answer, in one place, what all your agents did yesterday, what they cost, and which one a customer complained about? If the answer needs three dashboards and a Slack search, you don't have a fleet. You have a collection.

We build AI agent fleets that stay operable as they grow — shared evals, routing, and observability from the second agent onward — because the first agent proves the idea, and the platform is what lets the idea survive contact with the rest of the company.

Field note: A client came to us with four agents built by three teams in six months. Each one worked. Together, they were a mess: two of them called the same refund tool with different validation, one logged to a spreadsheet, and nobody could say what the fleet cost per resolved ticket. We spent two weeks not building a fifth agent but building the shared floor — one log format, one eval set, one routing table. The next agent shipped in nine days instead of six weeks, because the team started from the platform instead of from scratch. The lesson wasn't that platforms are exciting. It's that the second agent is where you pay for skipping them.

Related service Workflow automation

Building something worth shipping?

We take on a small number of AI product engagements. Tell us what you are building — we reply quickly.