The chronology of harness engineering shares a lot with that of human development. Here are some problems that both humans and agents experience, and their respective solutions:
| Problem | Human solution | Agent solution |
|---|---|---|
| Cannot effectively review one's own work | Peer review | Spin up another agent to review |
| Long problems are difficult | Notetaking | Agents write things down |
| Big problems are difficult | Organizations | Subagents |
| Hyperspecific workflows | Training | Skills |
| Some actions are dangerous | Security | Sandboxing |
| Getting lost in the weeds | Set a goal |
Set a /goal
|
| Slow, repetitive tasks | Scripting | Code mode |
It's hard not to say the human and agent solutions look alike. I think you can reasonably predict future harness improvements in the same way:
| Problem | Human solution | Predicted agent solution |
|---|---|---|
| Chain of command is slow | DMs | Talk across subagent trees |
| Stepping on toes | Meetings/Slack | Targeted and temporary agent chatrooms |
| Burnout | Time off | Being run with no task |
| Tribal knowledge discovery | Ask for help | Create and share skills with other agents |
| Lack of trust | Reputation, gossip | Create and share evaluations of agents with other agents |
| Good behavior is hard to define and always changing | Natural selection | Run experiments on altered copies of oneself, keeping the winners |
Some of these are extreme, but so were the ones in the first table. In fact in the time since I started thinking about this topic, several rows moved from the second table to the first.
I have to admit that even though it makes sense that this pattern would hold, I still find myself dumbfounded in the ways that it does. They somehow seem obvious in retrospect, but surprising in how minimal the right solution actually turned out to be—both for human development and harness engineering.