Radar
Where I'm finding leverage.
📌Pinned
Autoresearch: goal + eval, agents brute force the rest
If you can score an outcome, stop iterating by hand — let agents search the candidate space. Weco's tree-of-candidates harness and GEPA-style variant breeding turn tuning and optimization into a search problem. Half-ass the first version, or skip it entirely.
🔥Clutch
RLMs and recursive subagents
Don't have to manage context rot if you let the model divide and conquer large tasks recursively and aggregate the results. The natural successor to manually resetting your context window.
👀Watch
The Z/L continuum
The debate is no longer whether developers should write code — it's whether we should be reading it at all. Consensus is "yes, for now" if you care about drift; repo-understanding atrophy ranked #1 consequence of agentic AI in the AIE survey. But wire up enough adversarial review and xhigh frontier models, and maybe you don't.
👀Watch
Cerebras-class fast inference
OpenAI 5.6 sol at 750 t/sec on Cerebras — need to try this ASAP. Speed at that tier changes what interactive agent loops feel like.
👀Watch
Local frontier, one generation away
Gemma 3.5 4B gives you GPT-4o on your mobile phone, which is crazy. But DGX Spark and Mac Pro both have shortcomings — still one generation from Nirvana on local frontier power. You can have Nirvana today for $100k: the GB300 DGX Station.
👀Watch
Lights-out software factory
Cathedrals of higher-order loops, orchestration, and adversarial verification so humans sit at goal-setting altitude. The debate has moved from "should we write code" to "should we even read it." Not there yet, but tangible — not AGI hype.
👀Watch
GLM 5.2 heterogeneous routing
The talk of the World's Fair. Passively log production trajectories and test out of band which workflow steps can drop to GLM without quality loss — cheaper than building evals for every route.
🔥Clutch
Voice-first dev input
Wispr Flow and Handy give ~3x typing WPM, work anywhere, and handle filenames and technical terms. Pairs naturally with high-level goal-setting: talk to the factory instead of typing at it.
📌Pinned
Proactive SRE with full codebase context
Agents on cron review every block of telemetry, not just the spikes humans noticed. Loaded with the full codebase, they correlate errors and latency outliers to specific code paths and propose fixes. The "unknown unknowns" tier of SRE is finally cheap enough to staff.
📌Pinned
Suspended persistent microVMs
Firecracker-class VMs you can pause in milliseconds and resume just as fast, full state intact. Cheaper than always-on, more durable than serverless cold-starts. The unit of agent compute is shifting from request to session.
📌Pinned
Terminal computer use
Gemini 3 topped terminal bench with a simple harness, not an agent framework. Container runtime + read/edit/bash + agent SDK as inner loop. More durable than MCP, more practical than GUI computer use.
📌Pinned
Trendlines held
3-4 more OOMs of training compute on the horizon with Vera Rubin and Stargate. Scaling trendlines held steady through Gemini 3. UPDATE: METR's latest long-horizon agent results only reinforce the trajectory.
🔥Clutch
Progressive docs
Same shape as Claude skills: index up top, drill-down references, load context only when needed. Beats a 50-page wiki dumped into the window. Docs structured for agents are also clearer for humans.
🔥Clutch
Agents make OTel a dream
Instrumenting everything used to be a tax — now agents emit the spans, draft the dashboards, and write the APL queries. Observability finally pays for itself when reading and writing telemetry is free.
🔥Clutch
Generator-verifier gap
LLMs are better at verifying than generating. Exploit this with external verification loops (retry on failure) or internal ones (code execution + self-correction). Either way, verify rigorously.
🔥Clutch
Resetting context window
Don't use more than 50% of your context window. Use subagents for focused work. Start over when things go off track - don't try to steer back. Do research, store it, then leverage in fresh context.
🔥Clutch
Everyday max effort
Default to deep-thinking modes for everything, not just competition math. Mini research partners that work through problems methodically. Now available via API across frontier providers.
👀Watch
Claude agent skills
Progressive AGENTS.md with supporting resources. Easier to build than MCP, more flexible structure, avoids context bloat from heavy tool servers.