2026-05-28
May 28, 2026
Glue work didn't go away in 2026. The cheap kind got automated; the expensive kind got harder to see — and harder to staff.
Six years ago, Tanya Reilly stood on a LeadDev stage and named a problem most senior ICs knew but couldn't pitch: the connective labor that keeps programs alive — the unblocking, translating, hallway-following-up, "wait, has anyone told payments?" — was invisible on promotion packets and quietly killing the careers of the people who did it best. Her solution wasn't to stop doing glue; it was to see it clearly, decide which to keep, and refuse the rest.
Six years later, the question has inverted. AI agents now handle a meaningful slice of what used to be glue: status summaries, action-item extraction, doc cleanup, "who owns this?" lookups, mechanical follow-ups. Honeycomb has made doubling productivity through AI an explicit org goal. The 2025 DORA report has 90% of practitioners using AI in daily workflows, up 14 points in a year. Will Larson, in his April 27 essay on agents-as-scaffolding, was blunt: the recurring coordination tasks you used to do by hand are not coming back.
So the failure mode for senior TPMs and tech leaders has flipped. In 2019 you got stuck doing glue and never got promoted. In 2026 you get stuck doing glue that an agent could do, while the expensive glue — the kind agents can't touch — goes unstaffed and your org slowly stops working. The new craft is knowing the difference.
The mistake most senior TPMs make is treating glue work as one category. It isn't. It's at least three, and AI changes the math on each of them differently.
Mechanical glue. Status rollups. Standup notes. Meeting summaries. Doc syncs. "What did we decide last Tuesday?" lookups. Cross-posting the same update to four channels. This is what people meant by "glue" in 2019 when they meant "Tanya is doing the admin work no one will reward her for." In 2026, almost all of this is automatable. Fireflies-style transcript agents write your meeting summary. A scheduled Claude task pulls Linear, Jira, and Slack and assembles a weekly digest. The agent doesn't get tired, doesn't forget Wednesday, doesn't escalate at the wrong stakeholder. If you are still doing mechanical glue by hand in May 2026, you are paying a tax that your peers have stopped paying. Drop it. Build the agent. Tell your team you built it. That last step matters more than the first two.
Translation glue. Turning the staff DBA's three-paragraph rant into something the VP of Product can decide on. Reframing a security finding as a sequencing argument the program can absorb. Asking the question in the room that nobody else will ask because they don't have the cross-functional vocabulary. This is glue too — and it looks superficially like the mechanical kind because it produces an artifact (a doc, a summary, a recap). But the work isn't the artifact; the work is the judgment about what to compress, what to escalate, and what to discard. Charity Majors put it cleanly this spring: the new bottleneck is aggregation — synthesizing the parallel outputs of multiple agents (and multiple humans) into something coherent that someone can decide on. Agents can produce the first draft. They cannot reliably tell which draft is wrong, dangerous, or politically radioactive. Translation glue is appreciating in value, fast. Do not delegate it to an agent without a human read. Do not delegate it sideways out of false modesty.
Trust glue. Three engineers in a corner conflict you've been quietly mediating for two weeks. The PM who needs to be told her timeline is fiction, gently, before the all-hands. The director two levels up who will only commit to the migration if you specifically tell him the rollback plan is real. The skip-level you do because the EM is overloaded and a senior IC is about to quit. This is the work that doesn't render as an artifact at all. It's almost entirely judgment, relationship, and air-cover. Agents are not within five years of doing it. Will Larson's recent line — "judgment and creativity are all you need" — lands hardest here. Trust glue is the durable career capital of senior technical leaders in the AI era. It is also the easiest to lose if you spend your week doing mechanical glue out of habit, because you arrive at the hard conversations with no margin.
The senior TPM and tech-leader move for 2026 is straightforward, even if it isn't easy. Audit a week of your calendar against this taxonomy. Mechanical glue: how much, and why are you still doing it? Translation glue: are you doing enough of it, on the right surfaces, with the right read of who needs the artifact and who needs the conversation? Trust glue: which three relationships are you under-investing in because the cheap stuff is eating your week?
The 2019 advice was "stop doing the glue work that won't get you promoted." The 2026 advice is sharper: stop doing the glue work an agent can do, and start doing the glue work no agent will be allowed to touch. The first half is a tooling problem. The second half is a judgment problem. You only get paid for the second half.
Try this week. Open last week's calendar. For each meeting and each one-off task, label it M (mechanical), T (translation), or R (relationship/trust). Sum the hours. If M is more than 30% of your week, pick one M item and replace it with an agent or a template by Friday. If R is less than 20%, you have a quiet career-risk problem — block two skip-level coffees next week before someone else's calendar fills yours.
What it is. Tanya Reilly's framework, introduced in her 2019 LeadDev New York talk Being Glue and expanded in The Staff Engineer's Path (O'Reilly, 2022), for distinguishing technical-leadership glue (promotable, scope-expanding) from project-management glue (necessary but career-flat at IC ladders). It's the canonical reference for naming, valuing, and rationing the connective work that holds programs together.
When to use it. Quarterly career check-ins. End-of-promo-cycle retrospectives. Any time a senior IC or TPM on your team is "drowning but no one can name what they're drowning in." Also: planning your own week if you suspect your scope has drifted from leadership into administration.
How to run it:
When NOT to use it. During an active incident, a launch week, or the first 30 days in a new role — when all the glue is load-bearing and you genuinely don't yet know which strands you can safely cut. Run this framework on a normal week, not a crisis week.
Example: At a 1,500-engineer fintech, a Staff TPM ran this exercise in early 2026 and found she was spending 11 hours a week on mechanical glue — almost all of which her team's Claude-based weekly-digest agent could handle. She reinvested 6 of those hours into deliberate relationship work with three director-level peers and was named program owner for the company's largest 2026 migration in Q2. Her one-line takeaway: "The agent didn't save me time. It told me what my time was for."
2025 DORA Report: AI is the new baseline, and it amplifies what's already there — Google Cloud's annual study finds 90% AI adoption (+14 pts YoY) and 80% reporting productivity gains, but the headline for leaders is the warning: "AI doesn't fix a team; it amplifies what's already there." If your coordination is broken, agents make it broken faster.
The Pragmatic Engineer: AI Tooling for Software Engineers in 2026 — Gergely Orosz's spring survey: 55% of practitioners now regularly use AI agents, with staff+ engineers leading at 63.5%. The leadership read: your senior ICs are further ahead on agent tooling than you are. Catch up or stop having opinions on it.
Will Larson: Agents as scaffolding for recurring tasks (lethain.com, April 27, 2026) — Larson's argument: the right unit of AI leverage for a tech leader isn't "use AI in your IDE," it's what recurring coordination task can I retire this quarter? Frames the agentic shift as a scope-cleanup exercise, not a tooling one.
"The work I do that has the most impact rarely shows up in performance reviews. It shows up in the things that didn't break."
— Tanya Reilly, Being Glue, LeadDev New York 2019
Don't miss what's next. Subscribe to Critical Path: