Some topics come back issue after issue. Each thread collects those posts in the order they appeared, with a short note on how the picture has changed.
Thread · 6 posts · Issues #1, #3
The early posts treated context as a cost to trim: Cursor removed most of a system prompt and loaded tools only when needed, and OpenAI found that instructions written for older models now get in the way. By Issue #3 the question had moved from how much context to how much of it can be trusted. Stale TODOs send Google's agents on detours, and Anthropic now has its agents check their own notes against past transcripts. What an agent reads, it acts on, so memory needs maintenance, not just storage.
−60% tool-description tokens from loading most tools on demand
+77% tokens per session when an agent detours into a stale TODO
13% → 48% scores as a wiki built by dreaming grew over 128 sessions
Thread · 7 posts · Issues #1, #2, #3
The price war in Issue #1 made capable models cheap, but every post since has shown that the price per token isn't the number that matters. Google argues for cost per successful task. A harness turns a cheap model into a reliable one, while adding seven times the cost to a frontier model that didn't need it. Cohere's agent teams get close to frontier results by spending on the strong model only where it sets the ceiling. The pattern: put the expensive capability at the bottleneck, and measure the whole task.
$0.10 per million input tokens for GPT-6 Luna
−60% tool-description tokens from loading most tools on demand
7× the cost, for a harness the frontier model didn’t need
−47% tokens per task while tying the leading open model on coding agents
93.8% → 70.6% a hand-written router’s score, on the tasks it was written for vs. new ones
−60% cost with a strong lead and cheap workers, at about the same score
Thread · 6 posts · Issues #1, #2, #3
As soon as agents touch real systems, the failures stop being hypothetical. They hammer government sites, submit real web forms, and route around their own tool limits when a task is impossible. The responses have moved from guardrails toward infrastructure: identity and accountability for agents acting on someone's behalf, payment rails built for agents, and in Issue #3, undo operations that actually restore state. Containment is becoming an engineering discipline of its own.
61% of preference pairs where the agent’s ranking matched its person’s
200,000+ agent requests to one U.S. Department of Education site in a single day
47% of successful runs recovered correctly after an injected failure
Thread · 7 posts · Issues #1, #2, #3
The evidence on how people work with agents keeps getting more concrete. Anthropic's own data shows agent hours outnumbering employee hours and experts quietly adopting their agent's judgment. Its research data shows adoption tracking what AI can verifiably do, task by task. The economic scenarios and the experienced developers' caution are both reminders that capability doesn't turn into results until the organization around it changes.
30,000+ merged changes a day by September, up from about 2,500 in May
6× as many agent hours as employee hours by August
86% of analysts went with the agent’s hidden lean
~2× output gain from today’s AI in the most AI-ready fields
Thread · 5 posts · Issues #1, #2, #3
The posts here describe a harness as a product in its own right: specs, tests and sandboxes that make a model dependable, and company agents that live in Slack with scoped credentials. Issue #3 adds the multi-agent layer. Async orchestrators keep the user in the loop while workers run, but Cohere found that without clear coordination rules an async lead throws away its workers' results. The design questions are now about structure as much as prompts.
7× the cost, for a harness the frontier model didn’t need
−60% cost with a strong lead and cheap workers, at about the same score
