Field Notes/bryngtmagentsmcpthe-recorddreamforce

The Agent Your Agents Call

Dreamforce and UNBOUND open this week on the same idea: the agent lives inside the platform. Bryn is already the agent your agents call. Over MCP, a Claude, Cursor, or in-house agent can ask who is on your site from a fit account, why an account scored what it did, what ran on it and who approved, and hand Bryn a Play to install, with every answer drawn from the same record the UI shows. That is live today. Next, Bryn is learning to read a company the way a buyer does. Yours first.

Civic Team, Staff

Sep 14, 202614 min read

Title card: The agent your agents call. Below the title, two console frames side by side. The left frame is labelled your agent (Claude, Cursor, in-house) and shows three plain-language questions: which fit accounts hit pricing this week, why did acct_2210 score 86, what ran on it and who approved. The right frame is labelled Bryn, MCP, and shows the incoming call get_account_brief for acct_2210 and Bryn's reply lines: scored 86 against threshold 70, Fit 88, Intent 91, Timing 76, last run demo.follow-up on September 9 at 14:02 UTC, approved by op_3, record run_7d2e19, Logged. A single call line between the frames carries get_account_brief. Under both frames, one reply line in teal monospace: acct_2210, 86, pricing then comparison then repeat in 7 days. Illustrative, not a benchmark.

tl;dr

Dreamforce and UNBOUND open this week on one idea: the agent lives inside the platform. Bryn is already the agent your agents call. Over MCP, a Claude, Cursor, or in-house agent can ask who is on your site from a fit account this week, why an account scored what it scored, what ran on it and who approved, and hand Bryn a Play to install from your catalogue, and every answer comes back with its parts from the same record the UI shows. That is live today. Next, Bryn is learning to read a company the way a buyer does. Yours first.

Dreamforce runs September 15 to 17 in San Francisco. UNBOUND runs September 16 to 18 in Boston. The headline product at both will be the agent inside the platform, and the rest of the category has been building toward the same stage. HubSpot's Agent CLI, now in public beta, is a command-line interface built for agents rather than people to operate the CRM (HubSpot Developer Changelog, Jun 23, 2026). Outreach made its MCP server generally available with a Claude connector in February (Business Wire, Feb 24, 2026) and put its app in ChatGPT in June (Business Wire, Jun 3, 2026). 6sense piped intent into anyone's agent in August, which we wrote up as Intelligence Became an Ingredient.

All reasonable things to build. They leave one question open: where does the agent live when your pipeline runs across six tools and none of them is the platform? Our answer is somewhere the other agents can call.

The third surface

In August a founder described our architecture back to us on a feedback call, unprompted, for his own product: software now lives on three surfaces at once, the tool's UI, the agent inside the tool, and the agent outside it reaching in over MCP. Most of what gets announced this week is surface two. Bryn already has surface three.

"Callable" does a lot of work in launch posts, so here is what it means in practice. Bryn exposes an MCP server (docs). You copy a tenant-specific URL from the dashboard, your client signs in through Civic Auth, and whatever agent you already run, Claude Desktop, Cursor, VS Code, an internal bot, sees Bryn's tools. Your agent can then ask Bryn for:

the accounts on your site this week that fit your profile, ranked highest intent first, filtered by a signal or a minimum score (search_entities)
why an account scored what it scored, broken down by Fit, Intent, and Timing with the signals that contributed, from the same score data the dashboard shows (explain_score)
one account's brief: firmographics, last score, recent signals, recent Play runs (get_account_brief)
what ran on an account, who approved it, and when, from the append-only log, newest first (query_audit_log)
the Plays you have, the catalogue you can install from, and an install (list_plays, list_play_catalogue, install_play)

Two properties matter more than the list. The answers come back with their parts: the pattern by its steps, the score by its axes, the run with its approver. And nothing an agent does over MCP is off the record. The call is a line in the same log the UI reads, under the same membership that lets you into the workspace.

Ask Bryn what your agent would ask.

Four questions, in plain language. Each shows the tool the ask maps to and Bryn's reply, in its own record register, on one anonymized account.

Who is on my site from a fit account this week? (search_entities)

Three accounts above 70 carried pricing in the last 7 days. Top: acct_2210 (Anweledig Labs), 86, pricing → comparison → repeat (7d). Then acct_1187, 79, pricing ×2 (7d). Then acct_3402, 72, pricing → docs (7d). Ranked highest intent first, from the same list the dashboard shows. (illustrative, not a benchmark)

Why did acct_2210 score what it scored? (explain_score)

I scored acct_2210 at 86 on 2026-09-11 09:12 UTC, threshold 70. Fit 88: matches your ICP on industry and stack. Intent 91: pricing ×3, comparison, docs inside 7 days. Timing 76: returned within 7 days, twice this week. Same score data the dashboard shows. (illustrative, not a benchmark)

What ran on acct_2210, and who approved? (query_audit_log)

One run in 30 days. 2026-09-09 14:02 UTC I ran demo.follow-up against acct_2210 in Run mode: Slack alert to #growth, CRM task to the owner. The Play was approved by op_3 on 2026-08-19, once. Nothing blocked, nothing held. run_7d2e19. Logged. (illustrative, not a benchmark)

Install the pricing follow-up Play. (install_play)

Installed pricing.follow-up from your catalogue; the title is the catalogue's, not the caller's. Enabled in Approve mode, so each matching run holds for one click from you. First match: acct_2210, held for your approval at 09:14 UTC. Nothing has left. Logged. (illustrative, not a benchmark)

All values illustrative, not a benchmark. Account anonymized. Tool names are the live connector's; replies show the shape of what comes back, not a transcript.

Figure 1. Four questions an agent can put to Bryn today, the tool each maps to, and the reply in Bryn's record register (illustrative, not a benchmark).

What that changes

Pipeline review moves to wherever you already are. A growth owner opens a Claude session on Monday morning and asks which fit accounts carried pricing this week, why the top one scored what it did, and what ran on it. The answers arrive with their parts, and the decision that follows (approve the held run, or leave it) is the same decision the UI would have asked for.

An in-house agent can pull the account brief before a call and drop it into the prep. Here is the shape of what it gets back.

Figure 2. An account brief as get_account_brief returns it today: account, score by axis, the pattern by its parts, recent signals, last Play run, record id (illustrative, not a benchmark).

Governance carries over rather than thinning out. The agent signs in through the same Civic Auth membership that governs the UI, so it sees your workspace and nobody else's. When it asks Bryn to install a Play, it can only pick from your catalogue, the title comes from the catalogue and not from the caller, and the Play runs under the mode you set. In Approve mode, every matching run holds for a human click, whether a person or an agent asked for the Play. Every call lands in the log. We wrote in Bounded Autonomy Is the Spec that the boundary, the approval, and the record are what to demand from any agent before you hand it a channel. Putting your own agent in front of Bryn adds nothing to that list and removes nothing from it.

BRYN byCivicLabor Day offer ⬩ through September 17

Save Your Labor (Day)

Bryn watches your site, scores the account, runs the Play, and files the run. A free month of it, on any tier.

Timesheet ⬩ arbor.devPunched ⬩ Tue 2:02 PM

2:02:08 PMWatched a return to pricing, then the comparison page

2:02:09 PMScored the account 86

2:02:10 PMRan the pricing.follow-up Play into Slack and the CRM

2:02:11 PMFiled the run, replayable

Your time 0:00Bryn 0:03on the record

Claim your first month free

Reading a company the way a buyer does

That is step one, and it is live today on every tier. Step two is the only forward-looking sentence in this piece:

Bryn is learning to read a company the way a buyer does. Yours first.

Here is what the sentence means. A buyer evaluating a vendor does not read the homepage. They read the pricing page, the docs, the integrations list, the changelog, the careers page, and the comparison pages, and from those they form a view of what the company actually sells, how fast it ships, who it is hiring, and whether it fits. Brad wrote in July about what happens when the buyer sends an agent to do that read. This is the mirror image. Bryn is being taught to do the same read, for discovery, deep research, prep, and pre-sales.

It starts with your company because a brief about a prospect is only useful relative to what you sell. "They are hiring for growth" means one thing to a sales-tooling company and another to a payroll company. Before Bryn can say why a prospect fits, it has to know what fit means for you, from your own pages rather than from a form you filled in once.

What a buyer reads. What Bryn will read for.

Six pages a buyer evaluating a vendor actually reads. Pick one to see what a buyer infers from it today, and what Bryn will read it for.

What a buyer reads

What a buyer infers

What Bryn will read for

Pricing

Who the product is for, how it is sold (seats, usage, flat), where the ceiling is, and what "contact sales" is hiding.

Your buying unit and price ladder, so a prospect's plan and tier will map onto what you actually sell.

Docs

How real the product is, how much work adoption takes, and what the integration surface looks like.

Your capabilities as shipped, the setup path, and the vocabulary you use, so briefs will speak in your terms.

Integrations

Which stacks it fits and which it does not, and how it lands in a real environment.

The stacks you fit, so a prospect's stack will read as a fit signal or a gap.

Changelog

Velocity, what the team cares about, and whether the product is alive.

What shipped recently, so a brief will point at the newest thing a prospect's signals touch.

Careers

Where the company is investing, and what it is about to build.

Hiring as a timing signal on a prospect (a new growth hire is a window), and on your side, what you are building toward.

Comparisons

Who you think you compete with, what you claim to do better, and what you concede.

The alternatives you position against, so a prospect evaluating one of them will read as intent.

Illustrative. The right column describes a step in progress, in the future tense, with no date attached.

Figure 3. Six things a buyer reads about a company, what a buyer infers from each, and what Bryn will read them for (illustrative).

What does not change

The loop we laid out in July in What Bryn Actually Does: watch, score, run, log. The two modes, Run and Approve, and the kill switch behind them. The record. The read Bryn is learning becomes another input to the score and to the brief, not a new authority. It does not get to send anything, and it does not get to skip the Play. A better-informed score still clears the same threshold, still runs the same approved Play, and still writes the same line.

Then, now, next.

Three steps. Two are live. The third is in progress and carries no date.

Since launch (live). Bryn watches, scores, runs, logs. I watched pricing → comparison → repeat (7d) on acct_2210, scored 86, ran demo.follow-up, logged run_7d2e19.

Now (live). Callable by your agents over MCP. The same signals, scores, runs, and log, reachable by Claude, Cursor, or an in-house agent. Installs come from your catalogue; runs hold under Approve mode.

Next (in progress). Reads a company the way a buyer does, yours first. Pricing, docs, integrations, changelog, careers, comparisons, read into the score and the brief. No date.

Illustrative. Step 3 shows no date. Auto-advance is off by default and is not offered when your system asks for reduced motion.

Figure 4. What Bryn has done since launch, what your agents can do with it today, and the step in progress. No date on step 3 (illustrative).

Bryn is not another dashboard to watch. It is the governed execution layer that runs Plays through your stack.

Do this Monday

Point one agent you already run at Bryn's MCP endpoint. Claude Desktop, Cursor, an internal bot: the setup is a URL and a sign-in (docs). Then ask it three questions in plain language.

Who from a fit account was on our pricing page this week? Why did that account score what it scored? What ran on it, and who approved?

If the answers come back with their parts (the pattern, the score by axis, the record line with an approver on it), you have an agent your agents can call. If they come back as a summary, or do not come back at all, you have found the gap, and you found it before Q4 did.

Stop watching signals. Start running them.