The Agathon Framework
Agents crossed from demos into dependable work. What matters now is a discipline for choosing workflows, setting their autonomy, and knowing when to kill them.
Dr Colin Kelly
PhD in NLP (Cambridge)
Building production AI systems now
Most agent programmes stall late, because governance gets decided after the first incident instead of before the first deployment. Gartner's May 2026 guidance warns that applying uniform governance across agents of different autonomy levels leads to failure, and predicts around 40% of enterprises will demote or decommission autonomous agents by 2027 once the gaps surface in production.
The deeper cost of a bad first deployment is rarely the licence spend. A shallow rollout that changes nothing teaches the organisation that AI was overhyped, and that verdict is close to irreversible. The next proposal, the real one, has to fight uphill against your own institutional memory. The shallow deployment inoculates the organisation against the deep one, so the first deployment is worth designing as carefully as the tenth.
Pick the workflow first. Map where its decisions actually happen, then redesign the workflow around the agent. Bolting an agent onto a process shaped for humans preserves every coordination cost the agent was meant to remove.
The common failure is the broad rollout mistaken for adoption: a horizontal seat for the whole organisation, with licence count treated as evidence of progress. It produces a number that goes up while the value, which lives in re-engineering two or three high-stakes workflows, goes untouched.
The bar I use for whether a deployment counts: show me a workflow that no longer requires a meeting. Individual productivity gains are a perk that walks out the door with the employee. Workflow-level replacement, where the coordination itself disappeared and the shape of the team changed, is the asset.
Questions answered over your own documents, policies, and history, with citations.
Intake, extraction, classification, and first drafts of structured outputs.
Code, migrations, and testing under review. The most mature agent market there is.
Support queues, monitoring, scheduling, and escalation with clear rules.
Agents still burn money on open-ended judgement calls, domains too thin in data to check against, and any task you cannot write an evaluation for. If you cannot state what a wrong answer looks like, you cannot run the agent safely at any autonomy level.
This replaces the phased project plan. Every agent sits at a level, each level carries its own governance, and promotion is earned by passing an evaluation bar you set in advance. Never by a vendor roadmap, and never by a good demo.
Drafts, summaries, and answers. A human does the work and owns every output.
Governance: Data boundaries, so what enters and leaves the model is a decision, not an accident.
Promotion test: People keep using it without being told to.
The agent prepares the action; a human approves each one.
Governance: Full logging of proposal and decision, and an approval surface people actually use.
Promotion test: A measured agreement rate between agent proposals and human approvals, over a sample you defined before you started. Not a felt sense that it seems good.
The agent executes bounded actions; anything irreversible or outward-facing stays gated.
Governance: Hard limits written down, a termination path that has been rehearsed, and logs that can reconstruct any decision months later.
Promotion test: Error and rollback rates inside thresholds set before promotion was on the table.
Full loops inside an environment whose blast radius you control.
Governance: Purpose limits enforced by the platform rather than the prompt, and a kill switch someone has actually pulled in rehearsal.
Promotion test: Very few workflows need this level, and none should start here.
Most agent value arrives bought or adopted. Vertical agent products cover the common workload shapes well, and the tools you already run are growing agentic features you have already paid for. Building is justified in one case: the workflow encodes judgement or data that nobody sells, and owning that layer is worth the engineering it takes to keep.
The discipline is refusing to build what a model provider or vendor will ship next release. If your differentiator can appear in someone else's changelog, it was never a differentiator. I've written separately about how to tell what a vendor actually owns, and the same tests apply to the agents you are offered.
Five things that must exist around any agent before it earns autonomy
Agents break silently: the output stays fluent while the quality decays. A test set you trust beats any demo, and no promotion happens without one.
"Why did it do that?" must be answerable months later, with the context the agent saw at the time.
What the agent remembers, for whom, for how long, and how it gets purged is a design decision with legal consequences.
Connect tools through open protocols (MCP has become the default) so the integration outlives any single vendor.
Cost per completed task, not per seat. An agent that costs more per resolution than the humans it assists is a demo, not a deployment.
One more thing belongs in the operating layer, and almost every leadership team avoids it. The value of any agent deployment is gated by adoption, and adoption is gated by trust. The people who must change how they work are the ones whose roles are most exposed, and they are not stupid. Roll out agents while refusing to say what they mean for jobs, and people do the rational thing: use the tools shallowly and keep the workflow knowledge that would make them powerful, because that knowledge is their leverage.
So do the thing nobody does. Take five real roles and describe what each looks like eighteen months out with agents in the workflow. Who coordinates, who redeploys, who is genuinely redundant. Said out loud. You cannot extract deep workflow knowledge from people you are lying to, and that knowledge is precisely what the agent needs to be worth anything.
Choose one workflow. Map its decision points, redesign it around the agent, and write the kill criteria now, while nobody is attached to the outcome.
Real users on real work at proposal-level autonomy. Every action approved, everything logged.
Evidence builds against the bar set in week one. Nothing gets promoted without passing it.
Killing against pre-agreed criteria is a success outcome: it cost one quarter, and it preserved the organisation’s belief that the next deployment means something.
A platform purchased before a workflow was chosen
Seats deployed reported as the success metric
One governance policy applied to every agent regardless of autonomy
Autonomy promoted after a good demo rather than a passed evaluation
No tested way to stop an agent that is misbehaving
ROI slides produced before any workflow was redesigned
"The whole team has Copilot" presented as an agent strategy
In bounded shapes, yes: retrieval and answering over your own material, document pipelines, agentic engineering, and operational triage all pay reliably when the workflow is redesigned around them. They still burn money on open-ended judgement calls, thin-data domains, and anything you cannot write an evaluation for. The question has moved from whether agents work to which workflows, at what autonomy, under whose control.
Most agent value arrives bought (a vertical agent product) or adopted (agentic features inside tools you already run). Building is justified only where the workflow encodes judgement or data nobody sells. If a vendor can ship your differentiator in their next release, it was never yours.
Governance proportional to autonomy. An assistant that drafts needs data boundaries; an agent that acts needs hard written limits, logs that can reconstruct any decision, and a termination path someone has rehearsed. Gartner expects roughly 40% of enterprises to demote or decommission autonomous agents by 2027 because uniform, one-size-fits-all governance let gaps surface only in production.
One redesigned workflow can be live at proposal-level autonomy in about six weeks, with a promote, hold, or kill decision at day 90. Programmes measured in quarters and platforms usually stall before either milestone.
If the answer is not an immediate yes, that is the first thing to fix. No agent should act with real-world consequences until the termination path has been tested, not merely documented.
I help leadership teams pick the workflows, set the autonomy levels, and build the evaluation discipline this framework describes, as part of AI leadership advisory. For a quick read on where your organisation stands, the AI readiness assessment takes ten minutes.
Tell Me Which Workflow You Have in Mind