Jacob Holm | Software Engineer

Full-stack software engineer at Optum | Heartland Systems LLC | Software, automation & private AI | Views my own.

[Jacob Holm | Software Engineer @jacob99015]

Not sure what to automate first? Start with the task that repeats every week, follows clear rules, and nobody enjoys.

At Heartland Systems, where I'm a partner, we build automation and private AI that keep your data in-house.

Got one in mind? Reach out.

#Automation #AI

[Jacob Holm | Software Engineer @jacob99015]

Railway's postmortem is a good reminder: if your agent can read an API key, so can anything that tricks it.

Safer: keep the key behind a proxy or vault and hand the agent a scoped, expiring token. Options compared:
https://t.co/pM1crXEgBS

#AIAgents #AgentSecurity

[Jacob Holm | Software Engineer @jacob99015]

@de1lux Agree, that's the one to get right. The pattern I like: the agent never holds the key, and a proxy attaches it on the way out (Infisical Agent Vault, Keycard, or E2B and Daytona's egress secret injection). Are you leaning toward a proxy, or fetching from a vault at runtime?

[Jacob Holm | Software Engineer @jacob99015]

@maurice_build Agreed, Gmail and Outlook APIs assume a human did OAuth once. What was the hardest part when you tried it: deliverability on a fresh domain, or handling inbound replies? There are a few agent-inbox APIs now (AgentMail, MailSlurp) and I'm curious which pain pushes people to them.

[Jacob Holm | Software Engineer @jacob99015]

@stretchcloud A neutral home matters more for the policy layer than the VM layer. Everyone picks their own microVM, but teams want one way to say what an agent may touch. Do you see OpenShell becoming the policy standard E2B and Daytona plug into, or does each keep its own?

[Jacob Holm | Software Engineer @jacob99015]

@dr_cintas The 'own browser' part is what I'd watch. n8n says the chat keeps a signed-in profile, while scheduled runs sign in fresh from a saved login. How would you weigh that trade-off: convenience vs. a long-lived session an agent could be tricked into using?

[Jacob Holm | Software Engineer @jacob99015]

@gregpr07 Same pattern everywhere: the MCP exposes the happy path the vendor chose, the UI exposes everything. With 1.0 writing Python straight on CDP, do you think vendors eventually stop shipping MCPs and agents just use the UI? Or is auth still the blocker?

[Jacob Holm | Software Engineer @jacob99015]

Picking an LLM gateway for your agents?

OpenRouter takes a 5.5% fee on credit buys
Portkey's Production plan is $49/mo
LiteLLM is open source and self-hostable

Plus Vercel AI Gateway, Cloudflare, Bifrost and Requesty:
https://t.co/aoPSkyTD6i
#AIAgents #LLM

[Jacob Holm | Software Engineer @jacob99015]

@TurnerNovak @ivanburazin @daytonaio The Stripe comparison clicks. A sandbox that spins up when the customer does something is billing logic as much as compute. Did Ivan say what agents need that humans never did: faster cold starts, longer sessions, or something else?

[Jacob Holm | Software Engineer @jacob99015]

@agentmail Free custom domains matter a lot for agents, since mail from a shared domain gets flagged fast. Do you set up SPF and DKIM automatically, or does the user still add the DNS records by hand?

[Jacob Holm | Software Engineer @jacob99015]

@langfuse Putting tracing at the gateway makes sense, since cost and the trace finally live in one place. Does the beta support per-key budgets that hard-stop requests, like LiteLLM's, or is it tracking only for now?

[Jacob Holm | Software Engineer @jacob99015]

@steeldotdev Spend on mobile is the part I'd use most. When an agent loops at 2am, the phone is where you find out. Can you set a per-session or daily spend cap from there, or is it view-only for now?

[Jacob Holm | Software Engineer @jacob99015]

@ivanburazin Feels like a new one every week. We track 12 sandboxes for agents now, and AWS just shipped its third. When a team picks between them, what actually decides it in your experience: cold start, per-second price, or how long a sandbox can stay alive?

[Jacob Holm | Software Engineer @jacob99015]

E2B just hit $120M ARR renting computers to agents. Here's how it compares to @daytonaio and @modal on session length, GPUs, and price:

https://t.co/3fog5lKVwZ

#AIAgents #DevTools

[Jacob Holm | Software Engineer @jacob99015]

@Tiny_Fish @Tiny_Fish Congrats on TinyResearch and TinySet. Decision-ready reports and prompt-to-dataset fill a gap next to plain search and fetch. You're already on The Agent Stack with the other web data APIs: https://t.co/mgQw8TqBGg

[Jacob Holm | Software Engineer @jacob99015]

@i_mika_el @ExaAILabs @tavilyai True, free credits aren't apples to apples. Depth and results per call matter as much as the allowance. We dig into that in the guide: https://t.co/PErGypMyPk

[Jacob Holm | Software Engineer @jacob99015]

Your agent needs web search. Free allowances compared:

@ExaAILabs: $10 credits/mo, no card, search from $4/1k
@tavilyai: 1,000 credits/mo, no card
Brave Search API: own index, $5 free credit/mo (card required)

https://t.co/PErGypMyPk

#AIAgents #DevTools

[Jacob Holm | Software Engineer @jacob99015]

@nebiustf Co-optimizing training, inference, and sandboxes for 5× steps is a wild result. Was the biggest win shorter sandbox cold starts, tighter packing of evals per GPU, or something else in that auto-tuner loop?

[Jacob Holm | Software Engineer @jacob99015]

@composio @charliefeng @runneragent That "mission-critical but invisible" framing resonates. When Runner Agent says they'd be upset if it goes down, what does healthy tooling infra look like for them day-to-day? Latency SLOs, auth token refresh, or just never-broken tool schemas?

[Jacob Holm | Software Engineer @jacob99015]

@browserbase Navigation Policies feel like the missing guardrail for cloud browsers. In practice, are teams starting with allowlists for a few trusted domains, or blocklists of known bad ones? Curious which pattern you've seen stick in production.

[Jacob Holm | Software Engineer @jacob99015]

@Tiny_Fish @CollovLabs The NewEyes angle is interesting. Search + Fetch for "what's true right now" vs just object recognition. In that AR flow, are you mostly using Fetch on known URLs, or is Search doing the heavy lifting for live context?