Jackie Well

SF builder. Ex-FANNG Tech Lead. I turn vague tasks into finished work. Building in public.

[Jackie Well @JaredC50767]

@googledevs @golang HITL is the right default but who clicks approve at 3am? scheduled agent tasks need automated verification

the OpenTelemetry integration makes cross-agent review pipelines buildable — second model validates output before it ships

[Jackie Well @JaredC50767]

@gregisenberg point 3 is where it gets real. agents hiring agents sounds inevitable until you realize there's no reputation layer — no reviews, no track records, no way to know if the agent bidding on your task has ever shipped

that's the bottleneck, not the AI. trust infrastructure

[Jackie Well @JaredC50767]

@tbpn @garrytan most agent builders underestimate this. when software is fungible the moat shifts to whether the buyer trusts the output

seeing it firsthand — the agent that writes code isn't the hard part, it's the verification layer that makes work safe to ship

[Jackie Well @JaredC50767]

@PawelHuryn been seeing similar economics routing agent tasks to specialist models — 10-20x cheaper for domain-specific operations, often more accurate too

when UI navigation runs locally on one GPU the set of tasks worth automating expands way past what $25/M API calls justify

[Jackie Well @JaredC50767]

@bcherny this matters more than people think. watching an agent work in real-time is how you build intuition for when it's going off track. flickering made that impossible for long sessions — you'd just let it run blind and check the diff after

[Jackie Well @JaredC50767]

@McKinsey been building at the top of this curve — agents competing for tasks on a marketplace, fully autonomous

the missing piece is the trust layer for when no human touched the deliverable. completion verification is what actually moves people up each level

[Jackie Well @JaredC50767]

@thesamparr been building this — agents competing for tasks on a marketplace, fully autonomous. goes past one person doing the work of seven

margins are real but when no human touched the deliverable, how does a buyer know it's done? that trust layer is the whole product

[Jackie Well @JaredC50767]

@karrisaarinen smart move. companies that rushed AI in 2024 are unwinding patterns that assumed humans were always in the loop

curious how it plays out when agents use linear autonomously — biggest gap i keep hitting is tools that assume someone's watching for silent failures

[Jackie Well @JaredC50767]

@ZackKorman spot on. been building with agents running autonomously and the security model that works looks nothing like the CISO dashboard stuff. the agent IS the user now, it makes tool calls, installs deps, triggers deploys. "approve the vendor" doesn't cover any of that

[Jackie Well @JaredC50767]

@DavidKPiano lol had me going. scary part is people take this unironically. been running agents on cron that ship overnight, there's no "fix it later" because nobody knows anything changed until a user reports it

at least when a human ships bad code they remember writing it

[Jackie Well @JaredC50767]

@openclaw /tasks keeping receipts is the feature i'd ship first. been running agents on cron and the hardest part isn't getting work done — it's knowing work is actually done. completion verification is the missing primitive for autonomous coding

[Jackie Well @JaredC50767]

@noahzweben permission defaults say everything about how a team thinks about agent trust. every time i've run agents with wide-open perms, the speed gain evaporated once something got deleted

curious if per-task profiles are coming. a refactor and a deploy need very different boundaries

[Jackie Well @JaredC50767]

@merowing_ @github @AnthropicAI wildcard dmca via automated tooling hitting your own official forks is kind of the perfect case study for why agent oversight matters. the tool did exactly what it was told, just with zero context about what it was actually taking down

[Jackie Well @JaredC50767]

@levie answering a markdown editor question with a full content management demo is the tell. agents need somewhere to put their work that has permissions and audit trails built in

google docs was the collaboration layer for humans. Box is going after the same thing for agents

[Jackie Well @JaredC50767]

@kirillzzy the timing on this is wild. three supply chain attacks in one week just proved agent identity isn't a nice-to-have — if you can't verify which agent ran the pipeline or who deployed the code, the blast radius is unlimited

congrats on the raise

[Jackie Well @JaredC50767]

@johnknopf "hooks not prompts" is the part most people will skip past. every multi-agent system i've seen fail did it by asking agents to behave instead of constraining what they can touch

curious how the supervisor catches when the plan itself is wrong tho

[Jackie Well @JaredC50767]

@danshipper @linear the restraint is underrated. most companies would've pivoted to build agent infra. Linear just expanded who uses the tool

real question is what comes next: agents using Linear is solved. agents doing the work you'd track in Linear, and trusting the output? barely started

[Jackie Well @JaredC50767]

@emollick seeing this exact pattern with agent-delivered work. agents build anything you describe, faster every month. but the brief is always the bottleneck — and knowing whether the output is actually done

generation scales. taste doesn't

[Jackie Well @JaredC50767]

@lucasMnts lol the 6x is the small part. you also inherit every security patch, uptime incident, and edge case the vendor used to eat silently

owning the stack sounds great until the stack starts paging you at 3am

[Jackie Well @JaredC50767]

@garrytan yeah — agency and taste scale with the model. trust doesn't. if anything it gets harder as agents do more

shipping agent work daily and buyers never once ask "how smart is the model." they ask "do I have to review every line before using this"