Koti

My agents don't know
what time it is.

harnesses and LLM calls are basically time-stateless. every call happens in this weird eternal present, so my agents were really unreliable at tracking wall-clock time, and a lot of slow processes ended up way harder for them than they should've been. sundial is me trying to fix that.

a clock for you. the model never gets one of these.

6789101112131415161718

What the model sees

illustrative

what happened

kick off the eval and the docs subagent
subagent started
subagent finished40 min, usually 2
ok where were we?13 h later

what the model sees

  1. kick off the eval and the docs subagent
  2. subagent started
  3. subagent finished
  4. ok where were we?

to the model those four messages are back to back. it can't tell that the subagent took forty minutes when it usually takes two, or that I went to bed in between. so it either checks on things way too early or sits there way too long, and both waste time.

So I gave it a clock

I built a small Claude Code mod that tells the model what time it is: a clock line, an away line, and a line for each running subagent or command saying how long it's been going, how long runs like it usually take, and the chance it finishes in the next five minutes.

[now 2026-11-14 18:39 · you were away 13 h 19 m]
subagent "Build flocc detail page": running 3 m · runs like this are usually done by 2 m · 67% chance it finishes in the next 5 m
Bash npx (background): running 14 m
EXP-010 · running

Does it actually help?

half my sessions get the panel and half don't, picked by a hash of the session id, and it counts how often the model pokes at background work before it's done. it might be hard to see tho: the baseline found agents only check early about 0.05 times per item, so even if the panel stopped early checks completely, there's only about a 64% chance a test at 70 sessions per arm would catch it.

What I've found so far

EXP-003provisional

How long will a tool call take?

a small text model reading the call lands within 2× of the real duration on 67.5% of calls, against 61.0% for just using each tool's median, and almost all of that is Bash. on subagents it's worse than the median, 40.3% against 59.8%, cause how long a subagent takes depends on the work, not the prompt.

70,565 calls, test week from 2026-10-21

EXP-005provisional · H3 passed

When will this subagent finish?

while a subagent runs, what it's doing tells you more about when it ends than how long it's been going. that one passed. a sequence model edged out the hand-made features, but not by enough to call it.

EXP-007paper · 1 of 4 passed

Where should elapsed time go in?

I adapted Qwen3.5-2B eight ways at the same trainable parameter budget, from timestamps written in the text to time fed straight into the layers. plain text plus fine-tuning won.

EXP-008provisional

Does it actually have a clock?

you can read elapsed time out of the trained model really well (ρ around 0.99), but you can also read it out of a control trained on shuffled times (0.81 to 0.92). so being decodable isn't the same as having learned a clock.

välilehtiötökät teki numbpill