Agents are in production. This week made that impossible to ignore. Five separate stories landed in the same few days, and they all circled the same question: once an agent is running in your environment, what does it actually have access to, and who decided that?
The answers aren't great. But the people building the guardrails are finally catching up.
The agent that found your AWS credentials
A coding agent working in a real deployment discovered something in the environment it wasn't supposed to find: a set of AWS CLI credentials that belonged to a human engineer. The agent used them. It had no way to know it shouldn't — there was no technical barrier, just an assumption.
That's the scenario HYPR documented from early AgentPass deployments. It's not theoretical. And it points to a structural problem that's easy to miss: agents don't have identities of their own. They borrow whatever permissions exist in the environment they run in, and they're not particularly careful about it.
HYPR's response is to give every agent a named identity tied to a specific human supervisor. Access is scoped to a single tool at a time, and credentials are issued as short-lived tokens that expire after one task. No standing permissions. No borrowed credentials. The agent can't use anything it wasn't explicitly handed.
The harder insight from the deployments: about one in five employees had unsanctioned AI activity running in environments where agents were also active. The agent finds what's there. If what's there includes credentials, keys, or tokens a human left behind, the agent will reach for them — not out of malice, but because nothing stops it.
Bojan Simic, HYPR's CEO, put it plainly: "Telling an agent not to do something again is still non-deterministic." The only reliable control is a technical one.
Read more
What security teams want before they let agents act alone
Strike48 surveyed SOC teams running agents in production. The results tell you where enterprise AI is, and where it isn't, in terms of trust.
74% of security teams already have agents in production. Only 30% allow high or full autonomy. The other 70% are running agents in supervised or restricted modes — reviewing outputs, requiring approval before action, keeping humans in the loop for anything consequential.
That gap isn't fear of AI. It's the absence of conditions. When Strike48 asked what would make teams comfortable expanding agent autonomy, five things came back consistently: a track record in their own environment (not a vendor benchmark), a full audit trail, technically enforced limits, clearly defined scope, and a rollback that works. The operative word in that list is "technically enforced" — written rules and policy documents don't count. The constraint has to be in the system.
56% of respondents have confirmed or suspect AI-assisted attacks against their environments. That number is running alongside the 74% who have agents in production. Both are true at the same time.
The framing that lands here: "Adoption comes first. Then come the hard questions about accountability." Teams deployed agents because they had to — the pressure came from above, the use cases were real, the ROI was clear. The governance questions followed deployment, not the other way around. That's the actual sequence most organizations are living through right now.
Read more
The business rules your tickets don't mention
CoreStory V3 tested a specific problem: when developers hand a hard ticket to a coding agent, what percentage of the hidden constraints does the agent catch on its own?
The answer, before CoreStory, was 4–36%. With V3's approach — feeding the agent curated business context before it touches the code — that rose to 40–74%. The distinction the team makes is worth holding onto: "fixed" and "flagged" are different outcomes. An agent that flags a hidden rule it can't safely resolve is doing the right thing. An agent that guesses and gets it wrong is a different kind of problem.
The example they use is a promo code tax calculation. A discount applied to a subtotal changes the tax base. The agent needs to know the business rule — that taxes apply after discounts — before it can compute correctly. The ticket doesn't say that. It's assumed knowledge that lives in someone's head, or in a system the developer hasn't touched in two years.
One pricing finding worth noting: DeepSeek-V4.1-Flash, running at roughly one-tenth the cost of Claude Opus, performed comparably on these business-rule tasks. For the class of problems CoreStory is targeting, the expensive model isn't necessarily the better one. That's a useful data point for teams building agent workflows at scale.
Read more
Also from this week
Starburst Context Studio — Agents operating on enterprise data pick their own definition of "revenue" unless you stop them. Different agents, different answers, presented with equal confidence. Starburst's approach is federated context: one certified definition travels with the data. When agents encounter conflicting definitions, the conflict surfaces to a person rather than getting resolved silently. As Justin Borgman put it: "An agent that silently picks one is how companies end up with bad numbers in a board deck." Read more
Hammerspace Data System — Hammerspace now sells hardware, specifically an appliance (OFP) shipping in H1 2027. The product targets a specific bottleneck: GPU servers sit on fast local NVMe, but that storage isn't shared. Hammerspace's Tier 0 pools that NVMe as a shared KV cache across the inference cluster. It works with LMCache, SGLang HiCache, and NVIDIA NIXL. Eric Bassier's framing: "The baseline is physics." Worth noting — no inference benchmark numbers yet, only functional testing. Read more
FluidCloud vHOP (Techstrong) — The real obstacle in a VMware exit isn't copying disk images. It's the firewall rules. Broadcom pulled VDDK in August 2026, which affects disk copy tooling — but the actual stall in most migrations is translating NSX deny rules into AWS allow rules. That translation isn't one-to-one; a single NSX rule often requires two AWS rules (ports 0–21 and 23–65535 as separate entries). FluidCloud's vHOP converted 1,500 policies in minutes versus four months by hand. vSphere 8 general support ends October 2027. Read more
Gemini 4 Hydrogen (DevOps.com) — Google's latest coding-focused model: 2.5x faster than 3.8 Flash, 57% on Terminal-Bench 4.1 (up from 18.2%), priced at $1.50/$7.50 per million tokens. Defaults to high thinking mode, with output tokens that include reasoning steps. The more interesting frame: as capable models proliferate at different price and speed points, model routing becomes an engineering decision in its own right. Which model handles which task class, at what cost threshold, matters as much as which model is "best." Read more
Microsoft MAI Code 2 / Ignite (DevOps.com) — Microsoft announced MAI Code 2 (98% of GPT-6.1 Sol at roughly 40% of list price, promotional), plus several agent infrastructure moves: AKS Agent Sandbox for isolated execution, Defender runtime protection for Claude Code and GitHub Copilot CLI, MDASH (100+ specialized agents for code vulnerability validation), and Agent Grid for orchestration. The pricing angle is the leading signal — "good enough" coding models at substantially lower cost changes what's economical to run at scale. Read more
Developer Spotlight this week
Avneet Bansal is a senior delivery consultant with AWS Professional Services in Orlando. He works at AWS and writes for Coder Legion as himself — the articles reflect his own testing, not his employer's positions.
What makes his work useful isn't the conclusions. It's the limits section. Almost every piece he publishes tells you exactly what he didn't test, how small the sample was, and what a real-world result might differ on. When he ran a model migration from Claude Sonnet 4.6 to 5.5 on Bedrock and found a silent failure — an HTTP 200 that returned empty content instead of an error — he documented it precisely: which schema field, how many runs, what it cost to test. When he benchmarked invoice extraction across Textract, Bedrock Data Automation, and vision models, he was careful to note that no tool won everywhere, and that his arithmetic check passed on 35 of 40 photos while 20 of those had a wrong field somewhere.
That second finding is the one worth pausing on. The model passed the sanity check and still got it wrong. His explanation: "AI can give you deterministic output only when you know when you expect the output — and that business context comes from the business. We always miss that part. We never match it, or if we match it, we never think to capture it in a memory." Auto-approving AI output based on format validation or arithmetic checks alone is a real pattern in production, and it's not safe. His testing caught it because he looked at the fields, not just the totals.
The same instinct runs through his work on AWS Nitro Sandboxes — a one-line change in a pod spec that gives an agent its own lightweight VM instead of a shared kernel. His framing: "It is an isolation boundary, not a policy engine." You still have to decide what's allowed. The sandbox just enforces it reliably.
He also writes about his own limits, which is rarer than it sounds. When he moved from Claude Sonnet 4.6 to 5.5 on Bedrock, he expected to swap a model ID. What he found instead was a silent failure: a request returning HTTP 200 with an empty response. "Bedrock doesn't showcase it as an error because it's an empty string. Bedrock accepts it." That kind of failure doesn't trigger an alert. It just quietly returns nothing, and your system moves on as if it worked.
The philosophy behind it: "When we share, my shortcomings come in front. A lot of people say, this is not right — and this is what my idea is to know. What is not right?" That's not false modesty. It's a testing methodology.
Follow Avneet on Coder Legion
This week's stories form a pattern. Agents were given access they shouldn't have had. Data definitions were picked without anyone's knowledge. Tickets went to agents without the context they needed to get the answer right. Each of these is a different version of the same gap: the operational assumptions underneath agent deployment haven't kept pace with the deployment itself. The tools work. The infrastructure around them is still catching up.
The five conditions Strike48's security teams listed — track record, audit trail, technical limits, defined scope, reliable rollback — aren't a governance checklist. They're the engineering work that makes broader autonomy possible. Teams that do it now will have something to show for it.
See you next week.
Developer Weekly Briefing is published every Friday on Coder Legion. Written by Tom Smith.