Agents That Cannot Hurt You

What Softanza does differently from the 2026 agentic stack — for programmers, not AI specialists

This narration walks Softanza's answer and explains each idea before using it. You do not need to know what ReAct, a capability lattice or a trust posture is — that is the narration's job. Every code block below is real, and every output block is its actual output.

If you take one idea away, take this: the industry builds safe agents. Softanza builds a safe world and lets an ordinary agent loose inside it.

The premise everybody shares — and the one we do not

The mainstream stack assumes the agent is a language model in a loop. The model decides, so every safety mechanism exists to constrain a stochastic decision-maker: budget its context, judge its output with another model, cap its loops, log everything so you can reconstruct what it did. That is a large amount of engineering, and it is priced per decision, forever.

Softanza revokes that premise. Intelligence here is knowledge plus search plus optimization plus rules plus learning — running locally, explaining itself, costing nothing. A language model is one tier of the ladder, never its definition.

The consequence is not philosophical, it is structural. If the deciding part is deterministic, the safety problem stops being "constrain a guesser everywhere" and becomes "govern the one narrow door where a proposal turns into an effect." Almost everything below follows from that single move.

First idea: an agent is a file, and the file is judged before it runs

In Softanza an agent is not a class you subclass or a graph you wire. It is a declaration — a .pia file — dropped in a folder. And it is judged at load, not discovered at the first tick.

Here is an agent that reorders stock. It names a Haro function to do the work:

cBad = "pia: 2
name: restocker
kind: pi
coverage: reorders stock when it runs low
reversibility: compensable
schedule:
  timer: 50
skills:
  - name: reorder
    when: fact stock level low
    does: ring:PlaceTheOrder
    verify: fact stock level ordered
"
oBad = StzAgentDeclarationQ(cBad)
? "valid: " + oBad.IsValid()
? oBad.CiteFindings()

valid: 0
[pia-posture @ skills.reorder] this skill runs the Haro function 'PlaceTheOrder' and declares no execution posture. pia 2 requires 'posture: trusted | external | sandboxed' on any skill with a ring: clause -- the load gate says the function EXISTS; the posture says on what TERMS it may run (5.8).

Read the refusal carefully, because it is the house style: it names the rule (pia-posture), the place (skills.reorder), what is missing, and why the rule exists. Nothing was guessed and nothing ran.

Note what the file already had to say before anyone asked about safety: what it covers, and how reversible its work is. Add the missing line and it is admitted:

# the same file, plus one line:  posture: trusted
oGood = StzAgentDeclarationQ(cGood)
? "valid: " + oGood.IsValid()
? "format version: " + oGood.FormatVersion()
? "covers: " + oGood.CoverageStatement()
? "reversibility: " + oGood.ReversibilityClass()

valid: 1
format version: 2
covers: reorders stock when it runs low
reversibility: compensable

The industry equivalent of this file is a Python module plus a graph definition plus a tool schema plus a policy document. Here it is one artefact a non-programmer can read, and a court that refuses it out loud.

Second idea: two different questions, asked at different moments

Most stacks ask one safety question — may this action run? — at the moment the action runs. Softanza asks two, and the first one is asked when the answer is still free.

May this agent exist in the loop at all?

oGov = new stzGovernance("shop")
? "no coverage stated -> " + oGov.MayRegister("mystery", "", "reversible")
? "  " + oGov.Why()
? "no reversal declared -> " + oGov.MayRegister("mystery", "reorders stock", "")
? "  " + oGov.Why()
? "both stated -> " + oGov.MayRegister("mystery", "reorders stock", :Compensable)

no coverage stated -> 0
  AGENTLOOP-R4: agent 'mystery' declares no coverage statement. Registration refuses it. Say what this agent covers, in one sentence, before the loop will run it. Law 18: nothing schedules what nobody has stated the reach of.
no reversal declared -> 0
  AGENTLOOP-R5: agent 'mystery' declares no reversibility class. Registration refuses it. Declare one of: reversible | compensable | irreversible. Law 18: an agent whose reversal nobody stated cannot be scheduled by something that cannot undo it.
both stated -> 1

An agent nobody can describe cannot be scheduled. That question has no counterpart in the mainstream stack at all — there, an agent exists because someone constructed one.

And separately: may this actor perform this action, now? That is MayProceed, and it composes three things a stack usually keeps in three different places — what the actor may do, what it should do, and how dangerous the action is:

oG2 = new stzGovernance("billing")
oG2.DeclareRisk("send-invoice", 3)
oG2.GrantPermission("billing-agent", "send-invoice")
oG2.SetAuthority("billing-agent", :Delegated)
? "may it proceed? " + oG2.MayProceed("billing-agent", "send-invoice")
? "  " + oG2.Why()

may it proceed? 0
  refused: 'billing-agent' holds 'delegated' authority (level 2) but 'send-invoice' is risk tier 3 (SHOULD does not cover it)

Permission is not authority. The agent had the permission and was still refused, because its standing did not cover the danger of the act. Raise the standing and the same call is allowed, for a stated reason:

oG2.SetAuthority("billing-agent", :Autonomous)
? "after raising the authority: " + oG2.MayProceed("billing-agent", "send-invoice")
? "  " + oG2.Why()

after raising the authority: 1
  allowed: permission held AND 'autonomous' authority (level 3) covers risk tier 3

And an action nobody declared never proceeds at all:

? "may it do something nobody declared? " + oG2.MayProceed("billing-agent", "delete-ledger")
? "  " + oG2.Why()

may it do something nobody declared? 0
  refused: 'delete-ledger' has NO declared risk tier (undeclared actions never proceed)

That is fail-closed by construction. The usual failure — an agent doing something nobody thought to forbid — is not reachable here, because the default is refusal.

Third idea: how undoable an act is, not just how risky

Risk and reversibility are different axes, and conflating them is a common and expensive mistake. A low-risk action you cannot take back can deserve more ceremony than a high-risk one you can.

So Softanza asks who is allowed to do work of a given undoability. The rule is one sentence, and every door quotes it rather than paraphrasing:

? StzPostureReversibilityRefusal("sandboxed", "irreversible")
? "reversible: '" + StzPostureReversibilityRefusal("sandboxed", "reversible") + "'"

a 'sandboxed' posture does not cover 'irreversible' work -- the harder an act is to undo, the more trusted the code performing it must be (reversibility x posture, ruling 3.2)
reversible: '' (empty = allowed)

Trusted code may do anything; out-of-process code may do anything it can compensate for; LLM-composed code may only do what anyone can take back. In the mainstream stack this distinction lives in a paragraph of a design document, if it lives anywhere.

Fourth idea: the proposer cannot commit — by construction

Every serious 2026 guide recommends separating planning from execution: let the model propose, let a validated executor act. It is good advice, and it is advice — a pattern you implement, and can forget to implement.

In Softanza it is not a pattern. An LLM actor's effect-capability set is empty:

oLLM = LLMActor("summarizer")
? "capabilities: " + @@(oLLM.Kinds())
? "posture     : " + oLLM.Posture()
? "can cause effects: " + oLLM.IsEffectful()

capabilities: [ "inference" ]
posture     : sandboxed
can cause effects: 0

There is no configuration in which that actor commits anything. It can propose, and a deterministic gate disposes. Creativity is free; admission is governed.

Fifth idea: the agent works, and the world does not move

This is the part with no mainstream equivalent, and it is the heart of the thing.

Here is an ordinary agent whose job is to write a price list to disk. Nothing about it is special — it names a Haro function, exactly as before. The only new call is GiveWorkbench(), which hands it a virtual file system instead of reality:

oAg.GiveWorkbench()
? "cycle fired skills: " + oAg.Cycle()
? "agent believes it wrote them: " + oAg.MemoryQ().Fact("prices", "were", "written")
? "the real file exists: " + StzFileExists($cNarrOut)
? "the workbench holds it: " + oAg.WorkbenchQ().Exists($cNarrOut)

cycle fired skills: 1
agent believes it wrote them: 1
the real file exists: 0
the workbench holds it: 1

Look at those four lines. The agent ran. Its skill fired and verified. Its memory says the prices were written — and from the agent's point of view they were. The disk did not move. The work is real, complete and inspectable, and reality is untouched.

The function it ran was not written specially for this. It is the same code that would write to disk outside a workbench; it asks for the ambient bench and rehearses into it.

Sixth idea: the plan is the only thing that crosses

A tick's sole export toward reality is a plan — and a plan narrates itself, ranks its own risks, and can be argued with:

oPlan = oAg.GenerateUpdatePlan()
? oPlan.Narration()

Update plan (1 of 1 operations to commit):
  * 1. write 29 bytes to '.../_narr-fixture/prices.csv'

Hand it to the proposer and ask whether it may commit — the answer arrives before anything is attempted:

oPlan.SetExecutor(LLMActor("summarizer"))
aMay = oPlan.MayCommit()
? "may the LLM commit? " + aMay[1]
? "  " + aMay[2]

may the LLM commit? 0
  actor 'summarizer' cannot commit -- it lacks the 'effectful' capability (required by operation 1)

Now a committing actor, working inside a declared scope:

oPlan2.SetExecutor(PIActor("shop-committer"))
oScope = new stzCommitScope()
oScope.AllowUnder(cFx)
oPlan2.SetScope(oScope)
aRes = oPlan2.Execute()
? "committed: " + aRes[1][2] + ", skipped: " + aRes[2][2]
? "the real file exists NOW: " + StzFileExists($cNarrOut)
? "its content: " + StzFileRead($cNarrOut)
? "audit trail: " + @@(oPlan2.AuditTrail())

committed: 1, skipped: 0
the real file exists NOW: 1
its content: espresso,2.40
flat-white,3.10
audit trail: [ [ 1, "committed", "write_file", "shop-committer" ] ]

That is the moment reality changed — one place, by a named actor, with an audit line.

And the scope is not decoration. The same plan, committed under a scope that does not cover the path:

oScope2.AllowUnder("/somewhere/else")
aRes3 = oPlan3.Execute()
? "committed: " + aRes3[1][2] + ", skipped: " + aRes3[2][2]
? "reason: " + aLog[1][2] + " -- " + aLog[1][3]

committed: 0, skipped: 1
reason: REFUSED-BY-SCOPE -- path '.../_narr-fixture/prices.csv' is outside the allowed scope

A reviewer can also reject a single step with RejectOperation(n, :Because), re-validate the plan against a reality that may have moved since rehearsal, or read its ranked risks before deciding. The plan is a negotiation medium, not a verdict.

How this compares, honestly

Where the industry is genuinely ahead, and it matters:

• Connectors. OAuth'd Slack, Drive, Jira; hosted vector stores; streaming chat UIs. Reachable from here, not paved.
• Retrieval over open unstructured corpora. Hybrid search, re-ranking and freshness pipelines are mature there. Softanza's answer inverts the problem — curate knowledge and derive — which wins on a curated domain and loses on the open web.
• Frontier reasoning. That stack orchestrates the largest models per call.
• Ecosystem and hiring. There is a labour market for that stack and none for this one.

When to reach for which

Reach for the mainstream stack when the task is open-ended language work over material you cannot curate: summarising arbitrary documents, conversational products, research over the open web.

Reach for Softanza's agentic layer when actions have consequences and someone will eventually ask why did it do that — allocation, scheduling, back-office automation, regulated or audited work, anything that runs unattended, and anything that must keep working offline at zero marginal cost.

The two are not exclusive. The point of the capability lattice is precisely that a language model can sit inside a Softanza system as a proposer — creative, useful, and structurally unable to commit — while a deterministic gate decides what becomes real. That is the whole design in one sentence, and it is why an ordinary agent, loose in this world, cannot hurt you.

The mechanisms shown here are guarded by base/test/agentic/safeworld_narrated.ring (76 assertions), base/test/agentic/agentfile_narrated.ring (88) and base/test/governance/governance_narrated.ring (19). The doctrine is base/doc/design/Softaza and Agents That Cannot Hurt You.md; the ruling that bound the safe world to the agents is §3.2 of SOFTANZA_INTELLIGENCE_ARCHITECTURE.md. The industry positions summarised above are drawn from a 2026 agentic-AI engineering roadmap, paraphrased throughout.

Prouvé par

L'article nomme ces gardes de scénario de la bibliothèque :

• agentfile_narrated.ring(s'ouvre dans un nouvel onglet)
• governance_narrated.ring(s'ouvre dans un nouvel onglet)
• safeworld_narrated.ring(s'ouvre dans un nouvel onglet)

Citer

Mansour Ayouni. « Agents That Cannot Hurt You ». Softanza narrations, 2026. https://mayouni.github.io/stzsite/fr/narrations/stz-agents-that-cannot-hurt-you-narration.html

Non exécuté : son code touche aux fichiers, au réseau, à l'horloge ou au hasard. L'article est publié tel qu'il est écrit ; ses blocs sont des illustrations. Le code des articles est du code Haro. Là où un article a été écrit avant que la langue prenne son nom actuel, le nom a été changé et rien d'autre. Genre, domaine et série lus dans le nom du fichier, année tirée du premier commit : dérivés, jusqu'à ce que l'article porte son propre en-tête. Le fichier sur GitHub(s'ouvre dans un nouvel onglet)