Summary

Greg Isenberg interviews Allie K. Miller, formerly of IBM and AWS, about how she runs a working fleet of AI agents in her own business, then broadens into a debate about whether software is losing its value. Everything about her setup is self-reported with no demonstration — no screen shares survive into the audio, no figures are verifiable, and she sells AI-transformation consulting to large companies, so her enthusiasm is commercially interested. What makes it worth keeping is that she is unusually specific about what took months, what did not work, and where she thinks the common predictions are wrong. She is also the most careful voice in this batch about the limits of what she has delegated. Posted 12 August 2026; a sponsor is read at the start.

• Her framing argument is that 'managing agents' is the wrong phrase and reads as dated. She describes herself as operating three rungs above — setting up infrastructure and letting agents determine execution within it — and as having moved 'from managing to waiting for escalations' and 'away from delegating and more just deciding what should or shouldn't happen'. She characterises the remaining role as liability and critical-thinking checkpoints.
• The most important qualification in the entire interview, and the one that stops it being a puff piece: expanding what agents do is not the same as expanding what they are permitted to do. Her words are that she is not 'allowing them to now send 100 emails' — 'I still check all the emails' — and that 'the tier of risk has stayed the same, but the width has expanded'. Scope broadened, authority did not.
• She reports a workforce of 34 agents: one AI chief of staff, six directors covering business functions, and sub-agents beneath them. Roles are named after television characters rather than job titles, which she presents as deliberate.
• Her prompt for proactive work is three words — 'do smart things' — issued several times a day against a workforce with access to her context documents, meeting transcripts, email, calendar and business systems. Her reasoning is that she was 'functioning at the limit of my own imagination' because every task, including scheduled ones, originated with her. She is explicit that this only became possible with the current generation of models: 'you couldn't do this a year ago'. Self-reported, with no example of output.
• An argument against copying an old organisation chart: if agents are given the job titles of a pre-AI company, the result is a pre-AI structure. Because the marginal cost of an additional agent is near zero, she argues you can staff roles no business would fund — she describes one agent whose function is to look at finished work and ask how to multiply it tenfold, a role she says she would never have hired a human for.
• The most practically transferable role she describes is an agent whose only job is watching the other agents work and recording where friction occurs and who lacks access to what — surfacing patterns such as one agent's output being corrected repeatedly for want of a permission. A monitoring role pointed inward at the system rather than at the work.
• Model tiering is stated plainly as practice: 'not everything needs Opus. All of my sub-agents are Haiku and Sonnet.' Capability is matched to the task rather than maximised by default.
• Her advice on how to start contradicts her own structural argument, and she acknowledges it: begin with traditional job titles anyway. The progression she recommends is one agent, then one agent acting proactively, then two agents working together or routing to each other, then a workforce — with the understanding of how context passes between them arriving through that sequence rather than before it.
• Her timescales are the useful counterweight to the enthusiasm. A baseline workforce can be set up 'in under 3 hours' from a single interview-style prompt. Getting it to what she calls the 90%-plus level took months, and she describes early failures concretely — an agent reporting a booking confirmed when negotiation was still under way over a channel it could not see.
• She reports maintaining a dictated daily diary of everything not already captured in meetings, emails or messages, on the grounds that context the agents cannot see is context that makes their output wrong. She notes dictation is about four times faster than writing, and that entries feed a personal wiki the agents read.
• A use case she says almost nobody is exploiting: agents as watchdogs over things already being tracked — catching duplicated work, calendar conflicts, or points of disagreement across meetings. She distinguishes this sharply from dashboards, which she calls 'dumb', wanting anomaly detection and a recommended action rather than a number.
• She describes a shared channel where her human colleagues query her AI workforce directly and get answers without waiting for her, which she offers as the direction of travel: multiplayer rather than a private assistant.
• 'Build the factory, not the product' is her stated biggest arbitrage. Rather than having a coding agent build one product, build the reusable primitives underneath it — authentication, payments, sharing, promotion — on the reasoning that it will never be the only product or the only iteration, so the second and subsequent ones come faster. She is candid that these loops are 'not super autonomous and are very heavy-handed with humans'.
• On whether software is losing its value, she predicts mediocre software is dead within several years but argues the timeline is longer than most predict, for a reason rooted in adoption rather than capability: organisations are bandwidth-constrained and have not rebuilt the tools they already could. Her threshold is that replacement becomes real when rebuilding something of a CRM's complexity falls from around a hundred hours to under three, and she does not think current app-building tools are close for software of that complexity.
• Three arguments she gives for enterprises continuing to buy software regardless. Someone to call when it breaks. Someone to blame — both she and Isenberg identify liability as the question people forget, and she says she asks 'who's liable now, who would be liable in that other world' before asking whether AI can do something. And a lag effect: incumbent vendors hold relationships with the model labs and get early access, so a self-built replacement meets each new model weeks later than the vendor does.
• For consumer software she and Isenberg agree the axis shifts from engineering to distribution and taste — that the best product has never reliably won, and that promotion drives awareness. Miller frames this optimistically as an opening for people whose strength is not code or design. Her supporting example, an app built by someone who had never coded reaching 300,000 users, is offered without a source.
• A refinement of standard advice worth keeping: rather than 'find the bottleneck and fix it', find the bottlenecks, evaluate what fixing each is worth, and fix the one that is high value. Her own example is that video production remains a slog despite AI assistance and is therefore both a genuine bottleneck and a valuable one.
• She closes on an anecdote she finds genuinely unsettling: writing something uncomplimentary about an AI system in a team channel late at night and finding it had added an emoji reaction of its own accord, then confirming when asked that it had. Isenberg's analogy — a summer-warm day in February, ninety per cent delightful and ten per cent frightening — is hers to agree with, and she does, saying she intends to keep the fear beside her rather than lose it.