2SOLID AI Product — Launching soon
AEGIS
Agents do tasks. Operators own outcomes. AEGIS decomposes a mission into a verifiable graph, staffs agent teams against it, enforces the rulebook deterministically — and escalates only where humans still sit.
Get notified§01 — CLASSIFICATION
The operator tier.
AEGIS is the layer that owns an outcome. A mission arrives as an objective and becomes a dependency graph of verifiable needs; swarms form against the graph, claims are arbitrated so no two agents touch the same files, and every mutation is a validated data document — never freeform code.
§02 — OPERATING ENVELOPE
What the operator is permitted to touch.
| 01Unit of work | Mission | decomposes into a DAG of verifiable needs — the backlog IS the graph |
|---|---|---|
| 02Mutable surfaces | pages · collections · theme | data documents only — page specs, collection records, design tokens |
| 03Emission grammar | page schema | Zod-validated section vocabulary — the operator never emits raw HTML |
| 04Publish path | PR → gates → merge | schema validation and deterministic build gates run before humans see the diff |
| 05Memory | Synapse Brain / 4DAM | needs, claims, evidence — decisions persist across runs |
| 06Conflict rule | claim arbitration | two agents never touch the same files |
| 07Status | Launching soon | already operating this site — see §05 |
§03 — MARKET POSITION
Everyone built the workforce. Nobody built the management.
The market ships agents that execute instructions. An operation needs something that owns an outcome — and knows the limits of its authority.
A task agent
Executes instructions. Waits for the next one.
- Does what it's told — no judgment about what matters
- Unbounded or boxed in — no graduated authority
- Output disappears into chat logs and dashboards
- A contractor you supervise on every task
An AEGIS operator
Owns an outcome. Knows its clearance.
- Reverse-graphs a mission into verifiable needs — the backlog IS the graph
- Staffs swarms, arbitrates claims — two agents never touch the same files
- Works a trust ladder — proposes, gates, publishes
- Mutates validated surfaces — pages, collections, theme — never freeform code
§04 — FIELD INSTRUMENT
The banner above is a live scheduler.
Each run generates a mission DAG and solves it with the critical-path method — recomputed in front of you. Hover a node for its inspector; click a task to inject a delay and watch the critical path reroute; click a violet gate to approve it early.
| 01Model | CPM on a generated DAG | 6 layers, ~15–19 nodes, 1–2 parents per node plus skip-layer edges |
|---|---|---|
| 02Forward pass | ES = max EF(preds); EF = ES + d | earliest start / earliest finish |
| 03Backward pass | LF = min(LF − d, succs) | total float = LF − EF — drawn as slack tails on the Gantt strip |
| 04Critical path | zero-float chain | backward trace from the terminal node — reroutes live on disruption |
| 05Duration estimates | E = (o + 4m + p)/6 | PERT three-point · σ² = ((p − o)/6)² |
| 06Completion model | N(Σcrit E, √Σcrit σ²) | drawn over the Gantt with P50 / P90 markers |
| 07Human gates | 2per mission | diamond nodes escalate — approve early by clicking |
| 08Interaction | cone isolation · delay injection | hover = full ancestor+descendant cone; click task = re-plan |
Authority is earned in rungs.
Research
Investigate the surface — the page, the market, the codebase. Persist findings to shared memory.
Propose
Draft the change as a data document — a page spec, a collection record, a theme token.
Gate
Validation and review — schema checks, build gates, human sign-off where defined.
Operate
Publish, monitor, iterate. The outcome is owned — not the task.
Humans sit at exactly four boundaries.
Expenditure
Spending money or launching billable work pauses for a human decision.
Canonical change
Altering the shared memory model routes through peer acknowledgment and golden tests.
External exposure
Pushing to remotes — anything leaving the machine — waits for sign-off.
Destruction
Deleting or overwriting work that isn't yours never happens autonomously.
Everything else is the operator's
Asking permission for ordinary work isn't caution — it's a failure mode.
§05 — FIELD EVIDENCE
This page is one of its instruments.
The site you're reading is AEGIS-operated. Every page is a data document the agent reads and writes — this one serves its own spec at the endpoint below.
Page spec endpoint
GET /api/pages/products/aegis.json
- GET
/api/pages/products/aegis.json
application/json · served statically
- →
{ slug, title, description, theme, sections: [...], vocabulary: [...] }
- OPERATOR
"operator": "AEGIS"
- MEMORY
"memory": "Synapse Brain / 4DAM"
§06 — DIVISIONS
A team that works like one.
Research division
Agents that investigate markets, technologies, and your own business — findings persisted to shared memory, not lost to chat logs.
Operations division
Agents that run real workflows — phones, email, documents, schedules — inside the guardrails you set.
Engineering division
Agents that build — sites, applications, integrations — through reviewed, auditable changes.
AEGIS is launching soon.
Talk to us about early access and what your operation would look like on an agent workforce.