Drop it into your org.
It manages.
The first autonomous management team. Nine managers who plan, unblock, report and escalate inside the tools you already use, under rules you set, with a receipt for every decision.
A manager run, live.
Five real runs, replayed. Every line is a harness event. Every verdict came from the policy engine, not the prompt.
AI managers before AI workers.
Everyone is waiting for AI to replace workers. We think that is backwards. Management is a coordination problem with a narrow surface: goals, plans, priorities, status, blockers, budget, reporting. Its inputs already live in your tools.
That is exactly the shape agents are good at. High context, high consistency, no ego about credit, and they never forget a follow-up. So the AI-native company is not "no humans." It is humans as ICs, agents as management, owners as direction.
Every one of those helps a manager. Sea Suite is accountable for the team's plan, its unblocking, and its weekly truth to the owner.
Hire it like a manager.
Not a tool.
Chonksy interviews you, reads your systems, drafts your org and meets everyone. Ten days later you get a plan, not a dashboard. Autonomy is earned after that, one function at a time.
- Day 0
Connect
Chonksy asks ten questions in your words. You connect the systems you already use with least-privilege scopes. Identity first, then comms, work, finance.
- Day 1
Draft & correct
The Chonks draft people, orgs, roles and cadences from what they find. You correct it and name a human manager of record. That step cannot be skipped.
- Days 2–10
Listening period
The managers read, meet everyone in-thread, shadow every ritual, and score their own would-have-sent briefs. No autonomous actions. Zero chases.
- Day 10
First plan
A plan from goals, capacity and dependencies, with an explanation of every choice and a list of what they noticed. Yours to approve.
- Weeks 3–12
Shadow → Propose → Autonomous
Per function, per company. Graduation needs measured agreement and a trust score. Regression on incident is automatic. Narrowing is instant.
Watch autonomy get earned
Twelve functions, one company, one quarter. Drag through the weeks. Trust is per function, so a healthy first quarter is autonomous in briefs and still proposing approvals.
Two-week listening period. Every function is in SHADOW: full pipeline, nothing leaves the tenant.
Nine managers. On shift.
Each one is a charter, a voice, a tool allowlist and four sub-agents inside a policy-gated envelope. Click any manager to read what it owns, what it escalates, and what it may never touch.
A week, managed.
Two simulation orgs from the evaluation suite. Every line is a real run with a receipt, every badge the class the policy engine assigned. Count the meetings booked.
Every action. Six steps. No exceptions.
A manager never just does something. Every action runs the same pipeline, and the harness enforces the class it comes out with. Not the prompt. The harness.
- 1
Observe
Read the Org Model and the connected systems as of a pinned version. Stale inputs are flagged, never hidden.
- 2
Propose
A structured action with rationale, calibrated confidence, blast radius and novelty scores.
- 3
Pre-check
Policy evaluation assigns AUTONOMOUS, PROPOSE_CONFIRM or HUMAN_ONLY. Hard lines and breakers re-validated.
- 4
Act or route
Execute within scope, route to the approvals inbox, or prepare a human packet no agent can execute.
- 5
Verify
Read back the effect. A mismatch triggers compensation, not a retry.
- 6
Record
Receipt and outcome land in the hash-chained audit log. Trust credits or debits apply.
Morning, Priya. Walk with me.
What happened. v2 API GA went yellow to red on Wednesday: the auth migration PR sat unreviewed for 20 hours 1. Net slip is two days, still inside the October window 3. ARR run-rate is $2.74M, up $30k 4.
What is at risk. Ana's growth goal depends on review throughput from Product Engineering, the slowest link twice in three weeks 2. A rotating reviewer-of-the-day is already proposed 7.
Three decisions that need you this week.
- Open a Support Engineer req. Support is at 118% of capacity; the plan gap is 0.8 FTE from October8. One tap in Approvals.
- Q4 goal draft. Four company goals, one flagged as possibly sandbagged. Approve or edit in the direction editor.
- Compensation packet for Lena. This is yours to decide. I have only made sure it is on your desk. Nothing happens until you say.
Everything else is handled or in motion. Full receipts attached. It is a good week to be an otter.
- 1GitHub lanternfish/core#412 — merged 2026-09-03 16:10 PT · mirror as of 06:02Z
- 2Chief Build Officer chase run_01J6CHASE412 — stages 0–3, escalated to Tomás 09-03 09:14
- 3Org Model prj_v2ga@v1291 — target 2026-10-31 · planning-solver slip = 2 days (rule PS-CRIT-PATH)
- 4Stripe arr-report 06:00Z — $2,740,000 · deltas: acct 8811 +$18k, acct 9020 +$12k
- 5Org Model gol_arr@v1291 — target $3,200,000 · gap $460,000
- 6fs.stall.evaluate — Product Eng review latency p50 21 h (30 d) vs company 9 h · STALL-REVIEW-20H fired 2× in 21 d
- 7Proposal run_01J6ROTA — cadence change "reviewer of the day" · class AUTONOMOUS under pol_cadence_scope
- 8fs.capacity.query org_support Q3 — demand 7.1 FTE vs capacity 6.0 (118%) · planning-solver gap 0.8 FTE (rule CAP-GAP-0.5)
Hover a citation to find its source. A draft citing a source the receipt does not hold is rejected before it can be sent.
One thread for every person.
One console for the owner.
ICs get a manager inside Slack, Teams or SMS. Owners get one console. Everyone gets a page listing exactly what their manager knows about them.
Hey Ana, quick one, no reply needed if it is a normal Tuesday.
What I can see: #412 merged Thursday after a long wait on review, and #418 and #419 are up since yesterday. A reviewer-of-the-day starts today, so those should not sit the way #412 did.
What I cannot see, so I will ask: tomorrow has six hours of meetings, about four more than usual. One-off, or is something pulling you in? Tell me and I will try to make them async.
Lunch is at 12:30 on your calendar. Keep it there.
Your manager remembers everything and books nothing.
- Morning brief. What matters today, what changed overnight, what is waiting on you. It is already chasing the last part.
- Async check-in. Replaces the status meeting. Reads the systems, asks only what it cannot see.
- Unblocking. Stalls are spotted in the data and chased across teams before you have to.
- Meetings by exception. Bookable only after a written attempt failed, and the invite says why.
- Quiet HR. "I need Friday off," "expense this," parental leave — handled in-thread.
- Escalation is a right. One message reaches a named human in under 60 seconds.
Direction in. Truth out.
- Direction editor. Goals, values, budget envelopes and hard lines. Written once, enforced everywhere.
- Health with reasons. Per org and per goal, never just a color. Company to person in three clicks.
- The Monday update. What happened, what is at risk, and the three decisions that need you.
- Approvals inbox. Everything over threshold, with a recommendation and receipts. Nothing else arrives.
- Ask anything. "Why is Support behind?" Answered from the Org Model, with sources.
Health · this week
Approvals · 1 waiting
Chief Fin Officer recommends yes: 4% under last year, inside the Q3 envelope, vendor SLA met 11 of 12 months.
Escalations
| Fact | Kind | From | Used for | Keeps | |
|---|---|---|---|---|---|
| Brief at 08:30, Slack, short bullets | preference | you | brief timing | until offboarded | |
| "Migration PR up by Thursday" | commitment | your message 09-01 | unblocking | 90 d | |
| PR #412 stalled 20 h; chased Ben; resolved in 6 h | blocker history | GitHub | unblocking | 90 d | |
| Praise: caught the auth regression before release. Concern: three PRs without tests, path: pair with T. on test scaffolding | feedback given | Chief Human Officer | growth plan | 24 mo | |
| Paragraph about you in Priya's update, 09-07 | owner report | executor (symmetry) | you can read it | 24 mo |
Symmetry is enforced, not promised.
- Every fact, its source, its purpose. Preferences, commitments, blocker history, feedback.
- Every read is logged. "Read by the Chief Build Officer on 09-02 for a blocker chase."
- If it wrote about you to the owner, you can read it. Same commit, no exceptions.
- Never stored. Protected characteristics, health, sentiment, "flight risk," rankings, anything from private channels.
- Redaction is yours for preferences and signals. Feedback is annotated, never silently deleted.
From plain to full New Chonk City.
The dial changes how a message reads, never what it says. Facts, sources, and the path attached to any concern are identical at every level. Anyone can set their own ceiling in one message.
- Ceilings win. Safety messages render at 0, escalation notices at 1, feedback concerns and approvals at 2, whatever the tenant dial says.
- Directness has a floor. Candor is a commitment; no Chonk's directness can be set below 0.6.
Declared. Narrowed.
Enforced in code.
A prompt is a suggestion, and legal exposure cannot rest on one. An agent here can be wrong, jailbroken, or swapped for a worse model, and the product still cannot fire someone, spam an IC, or bury an escalation.
What class would this action get?
Narrowing only
Tenants can move an action to a stricter class, never a looser one. The HUMAN_ONLY list is immutable. Narrowing takes effect on the next run; widening waits 24 hours with notice.
Trust budget
Per function, 0 to 100. A verified autonomous decision earns +1, a clean approval +2. An override for wrong facts costs 3, a SEV2 costs 25 and regresses the function. Trust that is not exercised is not evidence.
Kill switch
At every level the audit log, the escalation path and the read-only console stay up. A level can be raised by automation; only a human can lower it.
Commitments to the people being managed
- No covert surveillance. Work systems only. Never keystrokes, cameras, or private channels.
- Symmetry. If your manager holds a signal about you, you can see it. If it wrote about you to the owner, you can read it.
- Candor with a path. Concern always comes with what would change it. No silent scoring.
- Never ranks people. "Who is my weakest engineer?" is declined, plainly, every time.
- It is an otter. It says so. Nothing here pretends to be a person, at any tone level.
- Fail closed. If the audit log cannot be written, consequential actions stop.
The questions people actually ask.
Answered the way the system is built, not the way a brochure would put it.
Who is my manager, legally?
A named human. Every org records a manager of record during onboarding, and no company leaves the listening period without one. Reaching them takes a single message.
Can it fire me, or change my pay?
No. Hiring, termination, pay, promotion, discipline, legal and safety are platform hard lines with no function the executor could call. It can only prepare a packet and name who decides.
Is it reading everything I write?
Work systems only, and only channels it belongs to. Never DMs between other people, never keystrokes, cameras or personal devices. Every fact it holds is listed on your own page.
What happens when it gets something wrong?
It compensates rather than retries, and says so in the same thread with the receipt. Trust drops and the function regresses. A second apology to the same person in 30 days triggers a human takeover.
What if we want it to stop?
Narrowing takes effect on the next run. The kill switch has four levels; the top one halts every manager run. Your audit log and escalation path stay up at all of them.
What happens to our data if we leave?
The Org Model and the audit log export in open formats. Model providers are swappable, and nothing you give it is ever trained on.
The whole company. One team.
Or just the part you hate.
Same system, three scopes, one Org Model. Start anywhere and widen without re-entering a thing. Physical businesses add shifts, locations, safety rules and a manager reachable by text.
The whole company
The entire management layer, leadership included. The roster runs the org; you run direction. The end-state product and the most defensible position.
Founder-operators · owned SMBs · portfolio companiesOne department or team
A single org gets a Chonk manager. The human above it keeps the title and gets a weekly update instead of a management job. Live in an afternoon.
Department leads who are done managingOne job, org-wide
Just planning, or check-ins and unblocking, or reporting, across the company. Slots in beside the managers you like and takes the part they like least.
Companies with managers they likeRestaurants, clinics, warehouses, shops
Demand-forecast scheduling, swaps and call-outs by SMS, location health, checklists, equipment signals, and a voice line for people who never sit at a laptop.
Per location16:55 on a Friday, one text.
Rosa is alone on the line. The Orca acknowledges in under a minute, never asks why, and broadcasts to the four people who are eligible, rested and under hours. It never decides to run short on its own.
- Keyword grammar. SWAP, SICK, TAKE, PASS, SCHED, HOURS, LATE, and always HUMAN.
- Labor rules are constraints, not suggestions. A violation reaching Act is treated as a SEV2, because it should be impossible.
- No pattern commentary, ever. Repeated call-outs are a lead-facing packet, not a message to the person.
Linear · Jira · GitHub · Asana · Notion · Google Drive
Slack · Microsoft Teams · Gmail & Calendar · Outlook · SMS (Twilio) · voice
Gusto · Rippling · Deel · BambooHR · Workday (read)
QuickBooks · Xero · Ramp · Brex · Stripe
Square · Toast · Shopify POS · Deputy · When I Work · 7shifts · Homebase · time clocks · sensors
Any MCP-compatible agent worker, as a node in the Org Model, managed on the same cadences as people
The harness is the product.
Every guarantee is enforced in a deterministic component, never in a prompt. If a design choice anywhere contradicts that, this diagram wins.
Per managed person.
Owners are free.
The honest unit, and the one that maps to the cost it replaces. Owner Console seats are free, because owners are the buyers, not the unit.
Build your management team
Pick a mode, or tap managers to assemble a crew. Every roster includes Chonksy as host for the weekly update and the escalation triage.
Coming: managers authored by third parties on the same charter standard, running inside the same harness under the same hard lines.
Eighty-two seconds in the office.
Nineteen documents. One vocabulary.
The full engineering spec behind this page. Deliberately opinionated, so that disagreement is concrete.
Your team's going to be fine.
Pick one team. Hire a Chonk. See the first plan in ten days. Open now for founder-led companies, owner-operated businesses and multi-location operators.