Entelligence Memory now powers the whole Entelligence platform.
• 𝗖𝗼𝗱𝗲 𝗥𝗲𝘃𝗶𝗲𝘄 checks every PR against your team's past incidents
• 𝗔𝘀𝗸 𝗘𝗹𝗹𝗶𝗲 answers why code was built and what broke after a deploy, in Slack, the web app or your terminal
• 𝗔𝗴𝗲𝗻𝘁 𝗜𝗻𝘀𝗶𝗴𝗵𝘁𝘀 shows how your team's coding agents actually work
Connect one repo and all three are set up.
Introducing @EntelligenceAI Memory: One shared memory for every coding agent on your team
- Sessions, PRs, tickets and incidents in one graph
- Agents have full memory access for all new sessions
- 53% fewer repeat errors
Available today for Claude Code, Codex and Cursor
STRAVA for your coding agents! 🏃
See how they run, fix their bad habits, pick up new skills and see whose agents are getting fitter alongside ur friends!
Completely new look and feel coming soon!
Teaching your coding agents new skills, reducing their token waste, optimizing md files, reviewing their design docs all now on autopilot
@harsh_logs rolling it out soon!
Two coding agents landed in AgentPlay this month. We didn't write either one.
Goose and Command Code came from community PRs. That's 10 agents in our open-source repo.
Six more are up for grabs: Qwen Code, Windsurf, Amp, Roo Code, Cline and Aider 👇
We looked at where agent tokens actually go.
The top 5% of sessions use 96% of them - 75% of the waste is compaction, rebuilding context after the window fills up.
No ceiling in sight, but most of the spend is a few sessions you can find.
The chart of token spend per employee still has no ceiling in sight. And the closer you are to AI right now the deeper the sense is that we’re actually still in the early part of the curve of token use and adoption.
This summer and fall’s model releases have continued to push the state of the art forward at a rate that already felt hard to maintain. As a result, for most productive use-cases the models are only now starting to actually work, and the feeling is that things will only accelerate from here.
Coding, in retrospect, was basically trash even 12 months ago. Now it’s finally starting to get good. And now that coding is actually working, you’re getting tool and computer use, which open up a wide variety of knowledge work domains that weren’t possible to throw agents at before.
Once you have agents that can code, use computers, leverage any tool out there, and understand a variety of domains, you can begin to throw them at a wide variety of knowledge work tasks that were hard to bring automation to across finance, life sciences, cyber, and more. So that brings agents to a broad set of work that we’re not really using them in yet at scale.
Then comes the new form factors of agents. Agent swarms and background agents are going to eat up vastly more tokens than anything we can imagine. You’ll have agents running in most enterprises workflows processing data, validating work, doing research, providing decision support, and so on. And we’ll have swarms that fan out in parallel in areas like cyber, life sciences, coding, and more.
So basically tokens are only going up from here.
More code isn't more shipped product.
We analyzed 1M+ PRs across 2,444 orgs and found only $0.18 of every AI coding dollar ships. Reverts grew 3.7x while PR volume grew 2.6x, and review is now the bottleneck.
research.entelligence.ai/
MORE AI CODE DID NOT TURN INTO MUCH MORE FINISHED SOFTWARE
Harvard researchers tracked 445 million work events for 720,542 workers at 655 firms, from January 2021 to August 2026.
After firms adopted AI coding agents, lines of code rose another 44%. Pull requests rose another 36%.
Finished work barely kept up. Resolved Jira issues rose about 14%, and the authors call that borderline significant. Human code review became the bottleneck.
On jobs, the paper says it "cannot attribute significant employment changes to AI."
When a company blames AI for cutting engineers, this study points to it being an excuse. For now.
We've started constantly auto optimizing our .md files
The context bloat that builds up over time otherwise can be actively hurting your agents
Almost need an agent monitoring your coding agents 24/7 and keeping them healthy
Everyone’s excited about spec driven dev/ factories but every company we talk to we keep seeing - claude writes the design doc, nobody really reviews it, claude builds it
claude still makes bad design calls. reviewing PRs for errors hardly matters when the entire design is off to begin with
Hiring a Research Engineer in SF at @EntelligenceAI!
Coding agents fail in repeatable ways: the same wrong assumption, the same missed test, the same review comments
We’re building the engine that makes them stop - a continuous cross harness/ cross model optimization engine
Fable 5.2 on max effort: $44.22 a task, scores 56
Same model on high: scores 58
You can pay more for your agent to think longer and get a worse answer.
Announcing Artificial Analysis Coding Agent Index v2.1, upgrading FrontierCode and Terminal-Bench. Claude Fable 5.2 and Claude Opus 5.5 share the lead at 58
The Coding Agent Index measures models in agent harnesses; v2.1 of the Index upgrades FrontierCode from 1.1 to 2.0 and Terminal-Bench from 4.0 to 4.1.
Upgraded:
➤ FrontierCode-AA 2.0 (replacing FrontierCode v1.1): our implementation of Cognition's FrontierCode 2.0, comprising 132 real-world coding tasks graded on whether a repository maintainer would merge the agent's change
➤ Terminal-Bench 4.1 (replacing Terminal-Bench 4.0): the same 66 tasks, scored with updated verifiers that improve reliability and resistance to reward hacking
Staying in the Index: SWE-Atlas-QnA-AA
Key results:
➤ Claude Fable 5.2 and Claude Opus 5.5 share the lead at 58: Claude Fable 5.2 scores 58 at both xhigh and high in Claude Code. Claude Opus 5.5 (max) scores 58 in Claude Code and, paired with GPT-6 Luna High, in Cognition's multi-model harness Devin Fusion CLI. Claude Fable 5.2 (max) scores 56 as more of its attempts hit task time limits. Claude Sonnet 5.5 (max, 56) in Claude Code and Gemini 4 Carbon (high, 55) in Antigravity CLI follow
➤ Index scores are lower in v2.1, driven by the upgrade to FrontierCode-AA 2.0: models score lower by 5 points on average. Codex with GPT-6.1 Terra (max) is the only configuration which didn’t fall between versions, sitting within 1 point of its prior score
➤ Cost per task varies ~30x among configurations scoring 48+: from $1.51 for Codex with GPT-6.1 Sol (max, 52) to $44.22 for Claude Code with Claude Fable 5.2 (max, 56)
come build with us!!
build the continuous cross-harness optimiz layer that pinpoints agent failures to the exact session step + causal path from pr reviews, aggregates recurring root cause errors and injects the learnings via the Entelligence mcp
hiring the world’s best!
Hiring a Research Engineer in SF at @EntelligenceAI!
Coding agents fail in repeatable ways: the same wrong assumption, the same missed test, the same review comments
We’re building the engine that makes them stop - a continuous cross harness/ cross model optimization engine
Hiring a Research Engineer in SF at @EntelligenceAI!
Coding agents fail in repeatable ways: the same wrong assumption, the same missed test, the same review comments
We’re building the engine that makes them stop - a continuous cross harness/ cross model optimization engine
Introducing Claude Meetings: add Claude to your team’s video call and it builds docs, whiteboards, and other artifacts with you while you’re still talking.
Starting to roll out in beta on Team and Enterprise plans.