Compacting code sessions is tricky in other ways (a summary loses a lot more important information) so defaulting to compacting isn’t great
I’d just stay in the safe zone before compaction is necessary, and using sub-agents to keep task tool use out of the main context. Coding harnesses are actually quite good at doing both of these, these days.
I’d just stay in the safe zone before compaction is necessary, and using sub-agents to keep task tool use out of the main context. Coding harnesses are actually quite good at doing both of these, these days.
1
0
0
32
Here's the full post. If you have thoughts or your own tips, I'd love to hear them. helloform.com/blog/harness-d…
1
0
1
113
On memory design: a "fact" from a web page looks identical to one the user told the agent. Record where each came from, and treat memories as facts, but not permission.
Hoping to write future posts on tools, guardrails, runtime architecture. Full post, with diagrams and code ↓
Hoping to write future posts on tools, guardrails, runtime architecture. Full post, with diagrams and code ↓
1
0
0
122
Think of conversation summaries as a letter to a successor who won't see the tool results.
Keep anything the user might need to point back to intact. A post-compaction example: "The agent presented several options" is useless when the next message is "let's do option B".
Keep anything the user might need to point back to intact. A post-compaction example: "The agent presented several options" is useless when the next message is "let's do option B".
1
0
0
62
On compaction: a cheaper summarizer model isn't always cheaper. Because it shares no cache with the conversation, it reads the whole history at full price.
Asking the main model, as one more message in the same context, if often better.
Asking the main model, as one more message in the same context, if often better.
1
0
0
64
Even sorted, editing memories busts the cache below it. Better: freeze sections, and append changes to the latest message ("since your instructions were written, this changed...").
Keep in mind the model sees now two versions of a fact, so it should be obvious which one wins.
Keep in mind the model sees now two versions of a fact, so it should be obvious which one wins.
1
0
1
90
Provider caching busts as soon as you change a token, so you should sort layers by how often they change: static at the top, per-turn at the bottom.
Classic mistake: current time timestamps in the system prompt, which breaks the cache on every single call. Ask me how I know.
Classic mistake: current time timestamps in the system prompt, which breaks the cache on every single call. Ask me how I know.
2
0
2
405
Agents resend the whole context on every call, often dozens of times per request. If you change one token near the top, you pay full price for everything under it.
Context layout is a cost decision. I wrote up a few months of harness design research. Highlights below 🧵
Context layout is a cost decision. I wrote up a few months of harness design research. Highlights below 🧵
5
3
14
1,964
how's that for job displacement
BREAKING: Microsoft, $MSFT, is offering buyouts to 60,000 workers as part of the company’s efforts to cut costs amid AI adoption, per Bloomberg.
Personnel on select teams in the US, Canada, India, and the UK at a vice president level and below will be eligible.
This reflects ~27% of Microsoft's total workforce.
1
0
3
1,021
@danshipper they saved that for later down the doc - can’t be a downer immediately
0
0
1
216
I use voice mode all the time. For the longest time there was no competition for ChatGPT voice mode. Claude is getting better but not quite there yet - unfortunately, because Opus 5.5 and Fable are my daily drivers.
I’ve had lots of success with @bot, though. The way the harness works means there are some quirks to it work around but it is the best at just getting things done, mostly on the back of the way I’ve structured agents and the delegation between them.
In short: for brainstorming, ChatGPT. For productivity on the move, Grok Bot. I assume Duet is going to change my views somewhat but this is where I stand today.
I’ve had lots of success with @bot, though. The way the harness works means there are some quirks to it work around but it is the best at just getting things done, mostly on the back of the way I’ve structured agents and the delegation between them.
In short: for brainstorming, ChatGPT. For productivity on the move, Grok Bot. I assume Duet is going to change my views somewhat but this is where I stand today.
🚙 had a few hours of driving today, so tried being productive w/ both Claude and ChatGPT voice mode
- drafted some follow-up emails
- researched some customer prospects
- practiced public speaking (!)
useful! but even hands free, it was distracting & impacted my driving 😞
2
0
5
1,344
@steveruizok 😂 can’t wait to double down on the performance reviews
0
0
1
135
@davemcclure @pitdesi Also worth noting that he has a great track record of making good predictions.
2
0
2
57
this morning, riding a bike
- "hey siri, play undertow by tool"
- "playing under-dough by tool"
- "play undertow by tool"
- "playing under-dough by tool"
- "fine. play lateralus by tool"
I guess I need to look into their bread-oriented discography
- "hey siri, play undertow by tool"
- "playing under-dough by tool"
- "play undertow by tool"
- "playing under-dough by tool"
- "fine. play lateralus by tool"
I guess I need to look into their bread-oriented discography
1
0
4
1,640
@dioscuri @alexolegimas @willdepue fantastic set of recommendations. Realizing I haven't read Know thyself yet. Fixing this today.
0
0
1
167
"wow, nobody warned us of all these risks"
but actually, lots of people did, quite often.
but actually, lots of people did, quite often.
I've been saying it a lot, but the damned if you do (talk about AI risk) damned if you don't (talk about AI risk) conundrum is going to get us in trouble.
If we stop disseminating important information because grifters and morons think it's all hype, we can't coordinate.
0
0
2
1,681