Notes
Notes on the field, and tests we ran first.
We read the primary sources so we know what is changing before we build on it. Every claim links to its source. Experiments are the tests we ran before building something, written up with the method, the data and the results.
34 of 34 articles
AI InsightsOct 24, 2026
Google's Harness²: the agent rewrites its own harness
Released in October 2026, Harness² lets an agent adjust its own tools and instructions for each task and keep what it learns. The model itself does not change.
AI InsightsOct 23, 2026
Three agent protocols, three jobs
ACP links editors to agents, A2A 1.0 links agents to each other, and AG-UI links agents to the people using them. All three are now governed by the Linux Foundation's Agentic AI Foundation.
AI InsightsOct 23, 2026
Agent Skills now belong to an open foundation
Anthropic gave the Agent Skills standard to the Agentic AI Foundation on 22 October 2026. It now sits beside MCP and AGENTS.md, with no single vendor in charge.
AI InsightsOct 23, 2026
Claude Code mods run outside the sandbox
A mod is a plugin whose code runs inside Claude Code itself, with your permissions. That makes mods powerful, and it means the sandbox does not protect you from them.
AI InsightsOct 23, 2026
Gemini 4 Argon's launch price is temporary
Gemini 4 Argon launched at $2 and $10 per million input and output tokens. Google's own footnote says the price will double to $4 and $20, the same as Claude Opus 5.5. Plan costs at the higher price.
AI InsightsOct 23, 2026
An 11-hour unattended agent run: 2 of 6 ideas shipped
Google left Gemini 4 Argon working alone overnight on an open-source project. It came back with six proposed changes, and the maintainer kept two.
AI InsightsOct 23, 2026
Briefing, 23 October 2026: Agent Skills go to an open foundation
Anthropic gave the Agent Skills standard to the Agentic AI Foundation. Maintainers from Anthropic, Google and OpenAI propose versioned releases and security guidance.
AI InsightsOct 22, 2026
Coding benchmarks no longer pick the best model
SWE-bench Verified is retired as a frontier test, and SWE-bench Pro has known flaws. Newer tests change often, and the useful measure is whether an agent succeeds every time, not once.
AI InsightsOct 22, 2026
How Claude Code keeps agents contained
Claude Code uses subagents to separate context and permissions, and operating-system sandboxes to enforce limits. An automatic classifier has replaced most permission prompts.
AI InsightsOct 22, 2026
Coding agents search code with grep, not embeddings
Most coding agents search files directly instead of using a vector index. Semantic search still helps in repositories with more than a thousand files, but agent design matters more than the search method.
AI InsightsOct 22, 2026
OpenAI, SpaceXAI and Google build the same agent factory
All three describe long-running coordinator agents that react to events and hand work to specialist agents. People still review the result, and guardrails are temporary and tracked centrally.
AI InsightsOct 22, 2026
Long context still needs a small working set
Anthropic and OpenAI now shrink conversation history on their servers. Even with million-token windows, agents work best when the active context stays small.
AI InsightsOct 22, 2026
What the labs say about context
Give the model the smallest set of information that's actually useful, watch how often your prompt cache gets reused, and let the agent look things up when it needs them rather than loading everything up front.
AI InsightsOct 22, 2026
Four ways to cut AI costs in 2026
Cached input now costs a tenth of the normal price or less, batch jobs cost half, and the two discounts stack. Sending simple work to smaller models saves more again. For reference, Claude Code runs about $13 per developer per day.
AI InsightsOct 22, 2026
What Cursor learned about running many agents
Swarms of equal agents failed. A planner, independent workers and a judge worked. Cursor's Projects adds a coordinator that never writes code and reacts to events.
AI InsightsOct 22, 2026
Small vision models now read PDFs best
For scientific PDFs, small specialised vision-language models now beat both traditional parsing pipelines and large general models. Claude and Gemini also read PDFs directly.
AI InsightsOct 22, 2026
Long-running agents now share a standard reliability layer
Agent steps are now logged so work can pick up after a crash, tools can pause for days waiting on a human's approval, and sandboxes plus classifiers decide what an agent is allowed to do.
AI InsightsOct 22, 2026
How teams test AI agents now
Teams start with 20 to 50 tasks taken from real failures and use exact checks before AI judges. When an AI does judge, it compares two answers side by side and never grades its own model.
AI InsightsOct 22, 2026
The labs now say: use our harness, not a framework
Anthropic, OpenAI and Google each ship their own agent runtime and recommend it over third-party frameworks. Frameworks remain for teams that need to stay vendor-neutral.
AI InsightsOct 22, 2026
How the big labs build research agents
OpenAI, Anthropic and Google use a coordinator that sends out parallel search workers, then a separate step that checks citations. None of them use a knowledge graph.
AI InsightsOct 22, 2026
Answering questions from papers is nearly solved. Citing them is not.
Simple agents over a good paper-search service already score about 85% on literature benchmarks. The open problems are citing sources accurately and combining them well.
AI InsightsOct 22, 2026
MCP drops sessions in its July 2026 spec
The new MCP spec removes sessions and the start-up handshake. Servers built on the 2025 spec are now on a path to retirement.
AI InsightsOct 22, 2026
MCP's two known problems: bloat and poisoned tools
Five MCP servers can use up about 55,000 tokens on tool descriptions before any real work starts. The other risk is tool descriptions written to mislead the model, and clients defend against that very unevenly.
AI InsightsOct 22, 2026
Agent memory is now plain files
Every major agent stores memory as files it reads and edits, with a short index loaded each time. Vendor memory benchmarks fail independent checks by 3 to 28 points.
AI InsightsOct 22, 2026
When to split work across agents
Separate agents work well for research, review and independent checking. Several agents writing the same code in parallel is still unproven.
AI InsightsOct 22, 2026
Search first, graphs only for multi-step questions
Combined keyword and vector search, with an agent that searches again when needed, is the baseline. Knowledge graphs only pay off on questions that need several linked facts.
AI InsightsOct 22, 2026
Agent systems now have two layers
One layer keeps the work running, with a permanent event log, a loop that can be thrown away and restarted, and a sandbox. The other decides what the model sees and which agent does what.
AI InsightsOct 22, 2026
Sandboxes are solved. The handoff after them is not.
Every major product now has a working sandbox, so this year's attacks go after the moment just after the agent finishes, when its output gets trusted. Scheduled triggers and webhooks still work differently from vendor to vendor.
AI InsightsOct 22, 2026
You can host your own copy of OpenAlex
The full research catalogue is free to download, about 745 GB compressed, updated quarterly. A copy limited to one subject is small. You lose daily updates and the hosted search.
AI InsightsOct 22, 2026
Two small files now carry agent instructions everywhere
SKILL.md and AGENTS.md now work across vendors: Codex reads the same skills standard, and Claude Code has read AGENTS.md since September 2026.
AI InsightsOct 22, 2026
Agent scaffolding goes stale with each new model
Newer models now do on their own what harnesses used to bolt on. Success rates barely change from one harness to another, but the cost per task can vary by up to 40 times.
AI InsightsOct 22, 2026
From "start simple" to "own the boundary"
In 2024 the labs told you to start with raw APIs. In 2026 the advice is to control the edges: keep checking separate from generating, put gates on tools, and strip out scaffolding as models improve.
AI InsightsOct 22, 2026
Briefing, 22 October 2026: the model matters more than the harness
Coding ability now comes from the model, not from the tools wrapped around it.
ExperimentsOct 19, 2026
Experiment: can a small model learn Oracle better than a large one?
We trained Qwen3-14B on public Oracle documentation and tested it against Claude Opus 4.8 on 379 questions neither had seen. Ours came out ahead on 289. Here is how we ran it, and what it does not show.