Software for readers
that pay by the token

May 23, 2026 · revised August 13, 2026

For forty years, everything software printed was written to be read by a person.

How we got here

It started with syslog. In the early 1980s, Eric Allman needed to see what Sendmail was doing, so he had it write lines to a file. There was no standard. It just spread, machine to machine, until it became the way software talked about itself. A log was a story a program told a human, one line at a time. Test runners, compilers, search tools, and shells all learned the same habit: print everything, let the reader skim. For a while, a human could keep up. Skimming was free.

Then the output grew past what any person could read, and we built tools to look harder: centralized logging, search, tracing, dashboards. Every one of those advances answered the same question. How does a human make sense of more output than a human can read?

The reader changed

Software is written and operated through agents now. Claude Code, Codex, whatever you point at the problem. And the agent inherits output built for eyes it does not have. It cannot glance. It cannot skim for free. Its context is finite and it pays for every token it reads. Watch a session and you see where the money goes: a test run that prints four thousand lines to report one failure, a search that returns a hundred files to use three, a log tail that buries the error under startup noise. The agent reads it all, once, and throws it away. Then it reads it again.

The output was built for the wrong reader.

Our bet

This is the start of an era, not an edge case.

The cost of running agents is dominated by reading, not reasoning.

Models get cheaper per token, and agents read more tokens. Sessions get longer, fleets get larger, and every tool call is billed input. When systems are watched and operated by other systems, the bottleneck is no longer human attention. It is whether the reader can afford what the world emits.

What we’re building

So we are building for that reader. Codag teaches agents to read large tool results efficiently: the evidence up front, cited and exact, with the full content retrievable the moment the agent needs it. Not a summary that loses the answer. Not a new workflow to learn. The same agent, doing the same work, reading like it can afford to.

We start with the tool results agents drown in first: logs, test output, search, file listings, documents, handoffs. It does not end there. The same gap runs through everything an agent has to understand through more output than it can hold. Anywhere that is true, we intend to be the layer underneath.

We serve agents first. That is the whole bet.

No one will miss watching an agent burn its window on a log tail, or paying twice for the same file read. That becomes history, the way reading raw syslog off a single box did. We are building the infrastructure that lets agents do more work per dollar, so the people running them can point them at bigger problems.

Michael ZhouFounder, Codag