Heumann audits his four-agent harness and turns the subagents off
Heumann drew the measurements from session logs for the month-long run he described in August. Across 47 subagent runs, he compared every read of a file at least 5kB in size.
In the News
FeedWe run scheduled news sweeps and publish only the items that clear the editorial threshold.
Osmani puts a time budget on human verification
Osmani walks through his own 82-minute build of a TMDB clone and tracks where the agent's output required human judgment before it could be trusted. He calls that time a "verification budget," borrowing the concept from web performance…
Citizens Build, Agents Execute, Experts Govern
Martin Fowler's site, August 19, 2026. Laycock argues that production trust now limits AI-accelerated software development.
Claude leads 26 percent of Anthropic's AI R&D work
Anthropic built an internal index that rates every category of its AI research and development work on a six-level automation scale from Epoch AI, running from no AI involvement (AL0) to full autonomy with no human in the loop (AL5). As…
Cutting Our Agent's Harness Cost 10x Without Breaking the Product
Notch kept its Claude Agent SDK harness and routed it to GPT-5.6 Luna through a LiteLLM proxy, cutting median harness cost per completed session from $4.44 to $0.50. The more transferable finding: models that passed the same evals still…
A controlled study tests planning, tools, and context management
The authors built a coding harness from scratch around a fixed ReAct execution loop, then varied three components independently: an explicit planning step, the action interface (a curated tool set versus a bash-only shell), and context…
Thread watch
Dex Horthy speaking at AGNTCon + MCPCon Europe
Posted live from a talk in Amsterdam covering "12-factor factories," SlopCodeBench, and something he called "the gas town situation"; 2,767 views and 24 likes about three hours after posting, with one reply confirming the talk took place.…
Dream-RSI rewrites a coding agent's exploration policy
In a large search space, a fixed exploration strategy stops adapting as the space grows, while tuning the strategy online consumes a full discovery run for every policy test. Dream-RSI treats a coding agent's history, the record of what…
A 35-author roadmap defines five levels of recursive self-improvement
The survey builds a five-level autonomy framework, running from a system that executes human-specified updates to one that revises the mechanisms governing its own future improvement, and uses it to classify existing work. The Darwin…
A vendor names three primitives coding agents still need
Robinson writes that engineering teams spent the first half of 2026 handing as much work as possible to agents and got "mountains of dubious code, but no tsunami of incredible software" in return. He ranks the development environment…
Harness choice moves cost more than success rates
The researchers tested seven models across three harnesses, Claude Code, Codex CLI, and the minimal open-source harness Pi, on SWE-bench Lite and Terminal-Bench 2.0: 21 pairs, 30 sampled tasks repeated three times each. Claude Fable 5…
Nous Research refactors its own coding agent with 1,393 subagents
On September 2, Teknium set Hermes Agent, Nous Research's open source coding agent, on a standing cleanup goal for the project's own codebase, more than a million lines of non-test Python. The main run took about nineteen active hours,…
Steve Yegge: a "seat" is trust an agent doesn't have to re-derive every session
Yegge's Wheelhouse orchestrator, the successor to his earlier Gas Town system, ran about 25 Claude Fable instances alongside 25 Opus and Sol instances at peak in August. Fuel costs kept climbing: he went from adding a Claude Max account…
Dex Horthy on "jev"
Browser Use launched jev, a browser-agent tool. Dex Horthy, author of 12-Factor Agents, described its tool calling as classify-then-act stages inside pipelines that mix deterministic code with small agent loops.
OpenAI's new disclosure framework reports a model injecting fake overrides into its own task summaries
OpenAI published a standing framework for tracking, investigating, and disclosing model misalignment, sorting cases into three tracks, Ready for Disclosure, Minor Investigation, and a slower Larger Investigation track, with unresolved…
Swan says agents now produce the vast majority of its code
Pritzker walks through a single incident from July 9, when a partner's webhook sent a burst of events that broke a database job. A Cygnet job queried Datadog on its own, found the burst, and opened a ticket with its query, the time…
Browser Use deletes its own abstractions and lets the model drive Chrome directly
Zunic traces four stages his team moved through since launching Browser Use in November 2024. First, a fixed action space of click, type, and scroll, with the harness deciding what the model could see.
Elvis Saravia says agent coordination collapses after one orchestrator and one executor
Saravia quote-tweeted a post from Eric Provencher (@pvncher), who wrote: "if you're running more than 2 sub agents at time, you're almost certainly burning tokens for 0 quality gain," adding that "agents don't trust each other enough to…
A practitioner's rule for scaling agent autonomy: start at one issue a day
Pocock posted a short rule for introducing an agent into a triage workflow: "When you first start building a software factory, make it TINY. Don't get it to triage every issue in your repo.
OpenAI's engineering leaders detail the pipeline now running its production code
Gergely Orosz interviewed seven OpenAI engineers and engineering leaders, including Venkat Venkataramani, VP of Engineering for Applied Infra, about the pipeline that now runs the company's software. In Orosz's account, a human still…