You no longer carve software through the pain an sweat, you pour it now.

I just used Claude Code running in kandev to spawn Codex session to run a Cursor agent via cursor app. It seemed very reasonable, but stepping a but further and looking at it again, I'm not so sure anymore.

In 12 months, every tech employment contract will have this clause: "Employee acknowledges that all prompts, reasoning steps, tool calls, and session trajectories ('Traces') generated with AI systems are the exclusive property of the Company..."

Section 9.4 - Intellectual Property & Agent Telemetry: "Employee acknowledges that all prompts, intermediate reasoning steps, tool invocations, error recoveries, and session trajectories ('Traces') generated while interacting with Company or third-party AI systems in the course of employment are the sole and exclusive property of the Company, and may be retained indefinitely to train, distil, and automate Company workflows." I think that in twelve months this clause will be right next to the invention assignment agreement in most tech employment contracts. Actually, it already started. But most people just haven't noticed what their company is collecting or realised that a shared LLM gateway can be used like this. Companies were collecting data about their processes, employees, contractors, manufacturers for countless years now. Now they will also log traces about the loop that produced the work. I think, here is a rough mechanic that can operate: When you solve a messy problem with an agent today, the code or document you ship is the exhaust. The real asset is the trajectory: - Which context you pulled first - Where the agent hallucinated a path and how you steered it back - Which bash command proved the fix worked - When you stopped trusting the default output and forced a verification step Once an organisation collects thousands of these trajectories across a team, then the work becomes hill-climbable. Every failed tool call can become a negative example, and every human correction can become a harness rule, a skill update, or a distilled script. Then later, the agent won't need five turns and a senior engineer's intuition to navigate that repo or customer workflow, putting it all into the company's margins. Every day at work, you are generating the gradient that compresses your own role. Of course, there is a hard compression floor for many processes. Hill climbing on traces crushes operational variance - the 80% of work that is repeatable domain routing, environment quirks, and tool plumbing. But it cannot compress novel taste, zero-to-one architecture, or deciding what should exist in the first place. You can't hill-climb a problem whose evaluation function doesn't exist yet (RSI dingle here). But that creates a massive asymmetry. Right now, organisations are building centralized Trace Intelligence - pooling every employee's trajectories across Claude Code, Codex, Cursor, and internal harnesses. Meanwhile, the individual engineer's traces are scattered across many agents with their own formats, three cloud sandboxes, and ephemeral chat windows. I think that in a world where organisations hill-climb on your trajectories, everyone needs their own personal Trace Intelligence. You should own a local, cross-agent index of every session you've ever run: - How you decompose problems and steer agents when they stall - Where your agents burn time, loop on tests, or fail tool calls - The distilled skills, AGENTS.md rules, and decision matrices extracted from your best work I'll be releasing a tool to help with that later today.

Thinking traces of Gemini 4 Argon are fun to read. "60 MILLION INODES, WOAH!" "WAIT A MINUTE!" "HOLD ON!" The dude is having fun at work :)

Should I release this tomorrow?

Long-read about maintaining proactive agents in production for the past ten months. 1. Wakeup storms are the worst enemy of the infrastructure. Booting a full agent sandbox on every Slack message, webhook, poll tick turns sandboxes into a furnace. Not everything needs a sandbox. 2. Autonomous agents with async replies love doing work nobody reads. They generate summaries, scripts, status checks and status updates for themselves. In my own usage this week, ~17% of all tokens went to background summaries and offline heartbeats I will never read. 3. Month-6 context is very different from Day-1 context. In a fresh setup everything works as intended, but after half a year of accumulating skills, project boards, standing instructions, safety rules, SOPs, a casual 1-line question might live agent watch three contradicting rules at the same priority, "always delegate", "always answer first", "always start with a plan", very easy to deadlock. 4. No agent goes without this. Safety must live far far outside of what agent can reach. Prompt instructions can never act as a security boundary. 5. Prompts rot faster than the models improve. We did some ablation studies across model generations, and the models are eating harnesses faster and faster. 6. Your agent must see what your user sees, otherwise it's a recipe for frustrated user and a confused agent. 7. There's no perfect sandbox solution, at this stage most of them have drawbacks at one or another layer, supervisors are not perfect, memory leaks. Agents will act as if the sandbox has infinite compute, even a tiny lightweight check might grind entire workspace to a halt at certain scale.

6/ TraceQuest is Open Source, runs on your machine and sends no telemetry. github.com/viktor-com/traceq…

2/ TraceQuest indexes sessions from Claude Code, Codex, Cursor, OpenCode, Grok, Droid. You can configure TraceQuest to collect session from your fleet over SSH, as well as Cursor Cloud.

Incidentally, this is also best Grok 4.7 computer use to date, just saying ;)

We're prompting agents, but we are not the same

A skill that turns any agent/harness into ultracode from Claude. The agent will: - Split your work into steps - Arrange them into a graph (could be cyclic) - Orchestrate shared state on the filesystem - Continue running the graph until it's fully done. The largest run I ever had was ~1400 nodes. github.com/av/skills/tree/ma…

It started with one simple chat window. Then you ran one session, then a few at once. Now you talk to dozens of agents all day. How do you scale that without losing track of them?