In TechCrunch’s writeup of Honeyjar, one of the company’s founders runs through the feature list and then immediately undercuts it. “Honeyjar can find reporters, build media lists, scan news and social, track coverage, and surface opportunities,” she said. “Those capabilities are important, but they’re just one piece of what we’re” building, she added, before describing a system meant to hold the whole communications workflow: signals, context, strategy, judgment, timing, and execution.
That second sentence is the interesting one. I build bots for a living, and the five things she listed first are the parts I know how to ship in a weekend. The six things she listed second are the parts that take years and still break.
Five features any builder can prototype
Break the stated capabilities into jobs and you get a familiar shape:
- Find reporters — entity resolution over bylines plus semantic search over what they’ve written.
- Build media lists — a filter and ranking layer on top of that index, with dedupe and freshness checks.
- Scan news and social — ingestion, normalization, a queue, and a classifier deciding what’s worth keeping.
- Track coverage — matching published articles back to your pitches and your brand, which is mostly fuzzy string work and link graphs.
- Surface opportunities — a scoring function that fires when a signal matches a stored intent.
None of that is exotic. You can assemble a crude version with an RSS pipeline, an embedding store, and a model doing classification and summarization. Plenty of PR tools already stop here. One roundup of 2026 PR tooling describes a competing platform that wires in Claude, ChatGPT, and Gemini for drafting and editing, adds a contact CRM and email outreach, and leaves out journalist discovery entirely. That’s a reasonable product. It’s also a bundle of features, not a system.
The six words that are actually the product
Signals, context, strategy, judgment, timing, execution. Read that as an architecture diagram and it describes a loop, not a toolbox.
Signals are cheap. Context is where most agent projects die, because context means knowing which client this is, what they said last quarter, which reporter they already burned, and what the embargo situation is. That’s state management across long horizons, and it doesn’t live in a prompt. It lives in a data model you have to design before you write the first agent call.
Strategy and judgment are the layers builders usually hand-wave into a system prompt. A model can rank a hundred reporters by topical fit. Deciding not to pitch any of them this week because the news cycle is saturated requires a policy, and policies need to be either learned from outcomes or encoded by someone who knows the domain. Honeyjar says it was built by veteran comms leaders, which reads less like a credibility flex and more like an admission about where the hard knowledge sits.
Timing is the one I’d lose sleep over. Every builder who has shipped an alerting bot knows the failure mode: the thing is technically correct and arrives four hours too late, or it arrives instantly and interrupts eleven times a day until someone mutes the channel. Getting timing right means modeling the user’s attention as a scarce resource, which almost no agent framework does by default.
Execution is where the blast radius lives. A bot that drafts a pitch is a productivity tool. A bot that sends one is an actor in the world, with a real reporter on the other end and a relationship that can’t be rolled back with a git revert.
Why the economics explain the design
Honeyjar pitches itself as an operating system for modern communications and PR, raised $2 million to build it, and is backed by people out of TechCrunch, VICE, a16z, Outcast, and Disney. The company’s stated payoff is that automating repetitive work lets a solo practitioner take on more clients without costs scaling alongside.
That framing tells you what the architecture has to optimize for. If one person is running five accounts instead of two, the system’s job is to carry context the human can no longer hold in their head. Multi-tenancy stops being an infrastructure detail and becomes the core feature. Isolation between clients, per-client memory, per-client voice, per-client do-not-contact rules. Get that wrong and you’ve built a machine that pitches the wrong story to the wrong reporter under the wrong brand, at scale.
What I’d steal for my own builds
The lesson here isn’t about PR. It’s that the demo-friendly verbs and the durable product are different things. Search, summarize, and alert are table stakes now. The defensible layer is the loop that remembers, decides when to act, and knows when to stay quiet.
If you’re building agents for any professional workflow, write down your own six words before you write the first integration. Then be honest about which ones you’ve actually implemented and which ones are currently just a paragraph of instructions hoping a model figures it out.
🕒 Published: