The harness we built to run our small dev shop with 2-3 people per product. 5 AI agents, one credit pool, one auth layer, 60 seconds of what the full loop looks like.
https://t.co/DrnHRsVu4c
GitHub Copilot switched to usage-based billing on June 1. Premium request units are gone; you now pay in "AI Credits" (1 credit = $0.01), metered by tokens — input, output, and cached — at each model's API rate.
The part that stings is agentic work gets billed twice. A Copilot pull-request review burns AI Credits for the tokens AND GitHub Actions minutes for the infrastructure running it. Some devs are posting projections of monthly bills jumping from $29 to $750. TechCrunch's headline: "What a joke."
Usage-based pricing itself is honest. The real problem is predictability and visibility: when the meter runs inside an agent loop, you learn what it cost after the fact.
That's the exact gap Codens built for. Budget caps are set per workflow and per org, the markup math is shown transparently, and Purple's run-level execution logs let you see where the tokens actually went, action by action. Claude runs API-direct, so the rate you pay is the rate, not a repackaged unit.
The agent era makes the meter matter. Caps and a visible ledger beat a surprise invoice.
Source (GitHub Blog): https://t.co/p8EpG17KVQ
https://t.co/zvNd4HPuaB
Interesting JetBrains research: awareness and actual usage of AI coding tools are pulling apart.
- GitHub Copilot: 76% awareness / 29% used at work, but growth has stalled
- Claude Code: 18% usage, but still growing fast
High awareness doesn't mean sustained use. Developers switch on whether the task actually gets done, not on brand. And the market's question has shifted from "are AI agents real?" to "which part of my company gets agentized first?"
Codens sits a little off to the side of that race. We're not competing on which autocomplete is smartest. We run the whole loop as a harness: PRD to tasks, implement, verify, PR, merge decision. The model or assistant underneath is swappable; what matters is whether the full loop closes and comes back in a form a human can review.
Not completion speed, but whether the work finishes. The adoption data is moving that way.
Article (JetBrains Research): https://t.co/qZ4UcSCOGI
https://t.co/zvNd4HPuaB
New in Codens: a safety guard that runs right before the AI pushes code.
It came out of a real incident. On one task the AI ran `git merge develop` to pull in test fixes, resolved the conflicts only partially, and pushed. Left alone, that would have landed 12,000+ lines of branch drift plus unresolved conflict markers straight onto the PR.
The guard has two layers:
1. Conflict-marker scan (always on, no config). If the committed tree still contains unresolved markers, the push is blocked.
2. Merge-source allowlist (configurable per workflow). Restrict which branches the AI is allowed to merge from. Set it right in the workflow editor; blank means no restriction.
The verify chain (implement → test → fix) and the human-gated / auto-merge choice are unchanged. This just adds one more layer before a human ever reviews. The more you let the AI do, the more these unglamorous guards pay off.
https://t.co/zvNd4HPuaB
Ran a quick benchmark. On a 50-memory dev corpus with 50 paraphrased (non-keyword) queries, 1 gold each:
Recall@1: 86%
Recall@3: 98%
MRR@10: 0.91
search latency: p50 20ms / p95 22ms
All local, nomic-embed-text + ChromaDB on an M5 Max. Retrieval quality is basically bounded by the embedding model; the MCP layer is thin. Small synthetic set, so treat as indicative, not SOTA. Repo: https://t.co/uYSTy7vfRO
Put yesterday's two stories side by side and the tide is clear.
- Anthropic filed for IPO and, for the first time, passed OpenAI in paid enterprise adoption — driven by Claude Code spreading from technical teams into finance, legal, and research.
- Microsoft Build 2026: its own Project Polaris replaces GPT-4 in Copilot, while Azure AI Foundry adds Claude as a first-class model alongside OpenAI.
The meta-signal: the "one model does everything" era is ending. Even Microsoft went multi-model — its own + Claude + OpenAI. The winning architecture isn't a single model, it's routing each task to the best one.
Codens was built this way from day one. Claude runs API-direct; work routes to self-hosted models or other lanes by use case. When a new model ships, one alias swap migrates every product. Don't bet on a model — bet on the structure that lets you swap models.
Article (Build 2026 recap): https://t.co/kzVXRRnThh
https://t.co/zvNd4HPuaB
A member of the Codens team just open-sourced a persistent memory MCP server for Claude Code. It's called Engram.
Re-explaining "I prefer Go" or "this project runs Postgres 16" every session quietly burns tokens. Engram remembers your preferences, decisions, and project context via semantic search, then pulls them back in the next session by meaning, not keyword.
And it's fully local. Embeddings run on Ollama, vectors live in ChromaDB, nothing leaves your machine. Zero embedding API cost too.
The unglamorous kind of tool that actually compounds. Written in Go, MIT licensed.
https://t.co/uYSTy7vfRO
Sharp call from Gartner: applying uniform governance across all AI agents will actually cause enterprise agent adoption to fail. They predict 40% of enterprises will demote or decommission autonomous agents by 2027 due to governance gaps found only after production incidents.
And 88% of agent pilots never reach production — not because models aren't smart enough, but because of governance, compliance, and data-residency walls.
What matters isn't the binary "trust the agent / don't." It's whether you can dial autonomy at the granularity of the task.
Codens designs this per workflow: auto-merge for low-risk work, human-gated merge for changes that touch production, and a verify chain that gates both. Not uniform — autonomy that scales with risk. In production, governance you can dial beats a smarter model.
Article (Gartner): https://t.co/40waRHdo3Q
https://t.co/zvNd4HPuaB
Today, June 1, GitHub Copilot moves from flat-rate to usage-based (AI Credits) billing. Pro $10 → $10 in credits, Business $19/user → $19 in credits, 1 credit = $0.01, consumed at per-model token rates.
The reason is obvious: across the industry, AI now writes ~41% of all code, and flat subscription can't carry that cost structure anymore. Microsoft just admitted it.
But flipping from flat to usage alone surfaces a new problem — ROI and governance. Gartner warns 40% of enterprises will demote or decommission autonomous agents by 2027 due to governance gaps.
Codens has run API-direct + per-workflow / per-org budget caps + transparent markup math from day one, solving both halves. Today's Copilot pricing change validates the position sideways.
Article: https://t.co/p8EpG17KVQ
https://t.co/zvNd4HPuaB
Strong Business Insider piece. Anthropic's head of growth Amol Avasare argues that for projects under 2 weeks of eng time, "the engineer is on the hook to effectively be the PM" — pointing toward the "product engineer" as the new hybrid role.
Right direction, but only half the picture. AI dissolves role boundaries in both directions: engineers becoming mini PMs, but also PMs / business becoming mini engineers. The second direction is actually higher leverage.
At Portament (which builds Codens), 3 business + 1 engineer shipped 5 products in 1.5 months, 1,000+ PRs. Engineer scarcity is faster to solve from the demand side than the supply side.
Article: https://t.co/fNIV93Sm60
https://t.co/zvNd4HPuaB
Three AI stories this week that connect into one picture:
- JPMorgan reclassified AI spend from "R&D experiment" to "core infrastructure". $19.8B budget, 2,000 dedicated staff.
- PwC is certifying 30,000 people on Claude to deploy at client engagements.
- Separately: one company reportedly burned $500M in a single month on Claude usage after failing to set employee spend limits (via @Polymarket).
AI is now infra, which means cost governance is now an infra responsibility too. Most teams are still bolting spend controls on after the fact.
Codens hits the Anthropic API directly, with per-workflow / per-org budget caps and transparent markup math built in from day one.
https://t.co/zvNd4HPuaB