Build AI agents that survive contact with production

The loop at the centre of an agent is not the hard part. A model, a list of tools, a while loop and a stopping condition is a weekend of work, and it will demo beautifully.

Everything after that demo is the project. This page is what we have learned building and running agents on our own business every day, including the parts that went wrong recently enough to still sting.

State is the project, and a store that only appends will forget

Agents get useful when they remember, and memory is where most designs quietly break. Adding to a store is easy. Deciding what leaves it, what merges with what, and what is allowed to be forgotten is the actual design work, and skipping it feels fine for months.

We measured our own on 30 August 2026 and it was worse than we expected: 1,175 behaviour rules accumulated, 24 of them actually reaching the working set, and 51 separate rules restating the same single lesson. Safety rules written one day were pushed out by near-duplicates the next.

The fix was not a bigger context window. It was consolidation with a similarity threshold and a real eviction policy, measured rather than assumed, which cut dead slots from 13.9 percent to 3.4 percent. If you cannot say what your agent forgets and why, it is forgetting something you needed.

Read every action back, because the tool will lie to you

On 30 August 2026 we found our own X integration had been logging successful posts that were never published. The API refused the call with error code 226 and the automation wrote its usual success line. Separately, three LinkedIn image posts logged as published are not on the profile, and fetching each one by id returns 404.

A response code is a statement about a request. It is not evidence about the world. Every action an agent takes should be read back from the destination by identifier, and anything that cannot be found afterwards is a failure regardless of what came back at the time.

This is one extra call per action and it is the single highest-value habit in agent engineering. It converts an entire class of invisible failures into ordinary loud ones.

The empty result is the failure nobody catches

Crashes get fixed because they are noisy. The failure that survives for weeks is the one that returns nothing and looks like a quiet day.

One of our lanes searched a community for four days with a filter that returned zero rows, because the plugin backing that filter was disabled in that category. No error, no exception, no alert, and a lane that should have been answering people simply did nothing while every dashboard stayed green.

So compare runs, not just responses. A search that returned results yesterday and nothing today is a claim that has to be defended, and an agent that treats an empty queue as normal will happily do nothing forever.

Identity and limits belong inside the agent

Anything an agent does on a platform carries an identity, and platforms notice when that identity is unstable. On 27 August 2026 our own browser automation was drawing a fresh device fingerprint on every launch, which looked to the platform like a new machine each time. It revoked a session cookie that still had 360 days of life left. The fix was to pin the fingerprint to a file and never regenerate it.

Limits belong in the same place. Every action our system takes on a platform stops at seventy percent of that account's live daily ceiling, enforced in the one function every caller must pass through, with the remainder waiting for tomorrow. A limit that lives in each caller is a limit that one new caller will forget.

The general rule: if a constraint exists to protect an asset you cannot easily replace, it goes on the shared path in code, not in a prompt and not in a runbook.

Check the output the way the user experiences it

A green build is not a check. On 26 August 2026 we discovered that the homepage of this site was serving search engines an empty container: no text, no links, no structured data, while taking 81 of the domain's 122 impressions for the fortnight. Every human visitor saw a normal website, so nothing looked broken and nothing was.

It was found by fetching the page the way a crawler does rather than by opening it in a browser. That is the same discipline an agent needs. Watch the video, open the document, call the endpoint and read the body, look at the destination rather than the log.

Frameworks help with the wiring and none of this. n8n, a graph library or a hosted agent platform will get you to the demo faster, and every failure on this page happens the same way inside all of them.