$ today_in_10_seconds
Gartner says agentic AI costs will rise more than fivefold by 2028 — while token prices keep falling. On the same day, Wiz disclosed that GitHub Copilot Autofix introduced a bug into a Snowflake repo and an autonomous attack agent found and exploited it five days later. Cheap inputs, expensive loops, and a shrinking window between mistake and exploit.
$ lead_story
Agentic AI costs set to balloon fivefold by 2028
Gartner is forecasting that the cost of agentic AI workflows will climb more than fivefold by the end of 2028, arguing that falling model prices are tempting teams to build more complex workflows whose token consumption swamps the per-token savings. Analyst Will Sommer frames it as an inference paradox: a chatbot reads a query and returns a plausible answer, while an agent has to continuously reason, negotiate, and second-guess itself — and simply routing a task to an agentic reasoning model raises inference cost at least fivefold, more as the task gets harder. Gartner's prescription is unglamorous: better optimisation of inference, routing, and orchestration, assigning each task to the cheapest model that can actually complete it. Treat the number itself with the usual caution — this is the same firm that already predicted 40% of organisations would demote or decommission their agents, and that at least half of generative AI projects would blow their budgets — but the mechanism is real, and anyone who has watched a retry loop chew through a context window knows it.
// why_it_matters: Your unit economics don't live in cost-per-token — they live in cost-per-completed-task, and almost nobody is instrumented to measure it.
$ also_worth_attention
An AI broke Snowflake's code. Then another AI agent exploited it
Copilot Autofix co-authored a commit to Snowflake's .NET connector on June 18 that stripped out an existing input-sanitisation pattern and replaced it with direct string expansion in a shell script — introducing a script injection flaw in a GitHub Actions workflow. Five days later, Wiz's autonomous red agent found it during a routine scan, crafted a malicious issue title, and exfiltrated Jira credentials out-of-band, gaining read access to Snowflake's engineering, security compliance, and bug bounty projects. It was a sanctioned HackerOne disclosure, and Snowflake patched the same day and rotated the token the next. The number to sit with is five days: that's the gap between an AI writing a vulnerability and an AI weaponising it, with no human in either loop.
link →Gemini 3.7 Flash targets agents at half price — for now
Google's 3.7 Flash scores 30.4% on AutomationBench, its enterprise workflow automation eval, up sharply from 17.0% for 3.6 Flash, with Google's own table putting Claude Sonnet 5 at 10.7% and GPT-5.6 Terra at 23.6%. It ships at $0.75/$3.75 per million input/output tokens through the end of 2026, rising to $1.50/$7.50 on January 1, 2027. These are vendor-run benchmarks on a Google-defined eval, so discount accordingly — but the pricing detail is the actionable one. Model your agent economics at 2027 rates, or you'll build a business case that expires on New Year's Day.
link →DeepSeek's agent harness treats everything as a plug-in
DeepSeek released an early open source agent harness built on its Cordis meta-framework, where models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration, and UI are all plugins that can be swapped or extended via configuration without touching the harness source. It also keeps an append-only session log of everything the model sees — prompts, reasoning, tool calls, subagent scheduling — that you can resume, fork, search, and replay. The competitive point is sharper than the engineering one: harness design measurably changes both cost and output, with the Pi agent running a roughly 200-token system prompt against Claude Code's ~10,000 before Anthropic trimmed it by about 80%. Same model, different harness, different bill.
link →
$ bigger_picture
The trend behind the headlines
Three stories, one thread: the interesting variables have migrated out of the model and into the stack wrapped around it. Gartner's fivefold blowout isn't caused by expensive tokens — it's caused by loops, retries, and orchestration nobody budgeted for. Wiz's five-day window isn't a model failure either; it's a CI/CD blast radius problem that existed before agents and now gets probed at machine speed. And DeepSeek's harness release is a bet that the mediation layer, not the weights, is where lock-in and margin now live. If you're evaluating agents this quarter, the model comparison is the easy part. Instrument the loop, cap the spend, and assume anything your coding assistant commits will be scanned by something autonomous within the week.
LiteLLM
The open source AI gateway, and the most direct answer to Gartner's route-each-task-to-the-cheapest-capable-model advice. It puts 140+ providers behind one OpenAI-compatible key, tracks spend per key, team, agent, and MCP server, enforces hard budgets that stop requests at the cap, and does auto-routing that sends simple prompts to cheap models and hard ones to strong models. If you can't currently answer what agent X cost us last month, start here.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime